use train_loss for sp test

* synthetic datasets for benchmarking and testing

* fix synthetic dataset parse from config and add tests

* use type=_synthetic

.github/CONTRIBUTING.md

@@ -68,7 +68,7 @@ You can skip certain CI checks by including specific keywords in your commit mes
 
 ### Code Style
 
-axolotl uses [{codestyle}]({URLofCodestyle}) as its code style guide. Please ensure that your code follows these guidelines.
+axolotl uses [Ruff](https://docs.astral.sh/ruff/) as its code style guide. Please ensure that your code follows these guidelines.
 
 Use the pre-commit linter to ensure that your code is formatted consistently.
 ```bash
@@ -83,6 +83,6 @@ Write clear and concise commit messages that briefly describe the changes made i
 
 - [GitHub Help](https://help.github.com/)
 - [GitHub Pull Request Documentation](https://docs.github.com/en/github/collaborating-with-issues-and-pull-requests)
-- [{codestyle}]({URLofCodestyle})
+- [Ruff](https://docs.astral.sh/ruff/)
 
 Thank you once again for your interest in contributing to axolotl. We look forward to collaborating with you and creating an even better project together!

.github/workflows/base.yml

@@ -15,6 +15,9 @@ on:
       - '.github/workflows/base.yml'
   workflow_dispatch:
 
+permissions:
+  contents: read
+
 jobs:
   build-base:
     if: ${{ github.repository_owner == 'axolotl-ai-cloud' && (github.event_name != 'pull_request' || !github.event.pull_request.draft) }}
@@ -124,7 +127,7 @@ jobs:
           images: |
             axolotlai/axolotl-base
       - name: Login to Docker Hub
-        uses: docker/login-action@v2
+        uses: docker/login-action@v3
         if: ${{ github.event_name != 'pull_request' && env.HAS_DOCKERHUB_CREDS == 'true' }}
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
@@ -132,7 +135,7 @@ jobs:
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v3
       - name: Build
-        uses: docker/build-push-action@v4
+        uses: docker/build-push-action@v5
         with:
           context: .
           file: ./docker/${{ matrix.dockerfile }}
@@ -173,6 +176,14 @@ jobs:
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
             dockerfile: "Dockerfile-uv-base"
             platforms: "linux/amd64,linux/arm64"
+          - cuda: "128"
+            cuda_version: 12.8.1
+            cudnn_version: ""
+            python_version: "3.12"
+            pytorch: 2.9.1
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+            dockerfile: "Dockerfile-uv-base"
+            platforms: "linux/amd64,linux/arm64"
           - cuda: "128"
             cuda_version: 12.8.1
             cudnn_version: ""
@@ -239,7 +250,7 @@ jobs:
           images: |
             axolotlai/axolotl-base-uv
       - name: Login to Docker Hub
-        uses: docker/login-action@v2
+        uses: docker/login-action@v3
         if: ${{ github.event_name != 'pull_request' && env.HAS_DOCKERHUB_CREDS == 'true' }}
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
@@ -247,7 +258,7 @@ jobs:
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v3
       - name: Build
-        uses: docker/build-push-action@v4
+        uses: docker/build-push-action@v5
         with:
           context: .
           file: ./docker/${{ matrix.dockerfile }}

.github/workflows/lint.yml

@@ -13,6 +13,9 @@ on:
        - ".pre-commit-config.yaml"
   workflow_dispatch:
 
+permissions:
+  contents: read
+
 jobs:
   pre-commit:
     name: pre-commit

.github/workflows/main.yml

@@ -8,6 +8,9 @@ on:
       - "v*"
   workflow_dispatch:
 
+permissions:
+  contents: read
+
 jobs:
   build-axolotl:
     if: ${{ ! contains(github.event.commits[0].message, '[skip docker]') && github.repository_owner == 'axolotl-ai-cloud' }}
@@ -110,6 +113,12 @@ jobs:
             pytorch: 2.9.1
             axolotl_extras:
             platforms: "linux/amd64,linux/arm64"
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.12"
+            pytorch: 2.9.1
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
             is_latest: true
           - cuda: 128
             cuda_version: 12.8.1
@@ -174,6 +183,7 @@ jobs:
     if: ${{ ! contains(github.event.commits[0].message, '[skip docker]') && github.repository_owner == 'axolotl-ai-cloud' }}
     # this job needs to be run on self-hosted GPU runners...
     strategy:
+      fail-fast: false
       matrix:
         include:
           - cuda: 128
@@ -259,6 +269,7 @@ jobs:
     if: ${{ ! contains(github.event.commits[0].message, '[skip docker]') && github.repository_owner == 'axolotl-ai-cloud' }}
     # this job needs to be run on self-hosted GPU runners...
     strategy:
+      fail-fast: false
       matrix:
         include:
           - cuda: 128
@@ -266,6 +277,12 @@ jobs:
             python_version: "3.11"
             pytorch: 2.9.1
             axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.12"
+            pytorch: 2.9.1
+            axolotl_extras:
             is_latest: true
             platforms: "linux/amd64,linux/arm64"
           - cuda: 128
@@ -326,6 +343,7 @@ jobs:
     if: ${{ ! contains(github.event.commits[0].message, '[skip docker]') && github.repository_owner == 'axolotl-ai-cloud' }}
     # this job needs to be run on self-hosted GPU runners...
     strategy:
+      fail-fast: false
       matrix:
         include:
           - cuda: 128

.github/workflows/multi-gpu-e2e.yml

@@ -20,6 +20,9 @@ concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}
   cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
 
+permissions:
+  contents: read
+
 env:
   MODAL_IMAGE_BUILDER_VERSION: "2025.06"
 
@@ -78,8 +81,9 @@ jobs:
           echo "AXOLOTL_EXTRAS=${{ matrix.axolotl_extras}}" >> $GITHUB_ENV
           echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
-          echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
           echo "E2E_DOCKERFILE=${{ matrix.dockerfile || 'Dockerfile.jinja'}}" >> $GITHUB_ENV
       - name: Run tests job on Modal
+        env:
+          CODECOV_TOKEN: ${{ secrets.CODECOV_TOKEN }}
         run: |
           modal run -m cicd.multigpu

.github/workflows/nightlies.yml

@@ -5,6 +5,9 @@ on:
   schedule:
     - cron: '0 0 * * *'  # Runs at 00:00 UTC every day
 
+permissions:
+  contents: read
+
 jobs:
   build-axolotl:
     if: ${{ ! contains(github.event.commits[0].message, '[skip docker]') && github.repository_owner == 'axolotl-ai-cloud' }}

.github/workflows/precommit-autoupdate.yml

@@ -5,6 +5,8 @@ on:
     - cron: '0 0 1 * *'  # Run monthly
   workflow_dispatch:  # Manual kickoff
 
+permissions: {}
+
 jobs:
   auto-update:
     runs-on: ubuntu-latest

.github/workflows/pypi.yml

@@ -3,9 +3,11 @@ name: publish pypi
 on:
   push:
     tags:
-      - 'v*'
+      - "v*"
   workflow_dispatch:
 
+permissions: {}
+
 jobs:
   setup_release:
     name: Create Release
@@ -28,7 +30,8 @@ jobs:
       name: pypi
       url: https://pypi.org/p/axolotl
     permissions:
-      id-token: write  # IMPORTANT: this permission is mandatory for trusted publishing
+      contents: read
+      id-token: write # IMPORTANT: this permission is mandatory for trusted publishing
     steps:
       - name: Check out repository code
         uses: actions/checkout@v4
@@ -46,7 +49,7 @@ jobs:
 
       - name: Extract tag name
         id: tag
-        run: echo ::set-output name=TAG_NAME::$(echo $GITHUB_REF | cut -d / -f 3)
+        run: echo "TAG_NAME=$(echo $GITHUB_REF | cut -d / -f 3)" >> "$GITHUB_OUTPUT"
 
       - name: Update version in VERSION file
         run: |

.github/workflows/tests-nightly.yml

@@ -8,6 +8,9 @@ on:
     paths:
       - '.github/workflows/tests-nightly.yml'
 
+permissions:
+  contents: read
+
 jobs:
   pre-commit:
     name: pre-commit
@@ -156,8 +159,9 @@ jobs:
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
           echo "E2E_DOCKERFILE=${{ matrix.dockerfile || 'Dockerfile.jinja'}}" >> $GITHUB_ENV
           echo "NIGHTLY_BUILD=${{ matrix.nightly_build }}" >> $GITHUB_ENV
-          echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
       - name: Run tests job on Modal
+        env:
+          CODECOV_TOKEN: ${{ secrets.CODECOV_TOKEN }}
         run: |
           modal run cicd.e2e_tests
   docker-e2e-multigpu-tests:
@@ -198,7 +202,8 @@ jobs:
           echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
           echo "NIGHTLY_BUILD=${{ matrix.nightly_build }}" >> $GITHUB_ENV
-          echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
       - name: Run tests job on Modal
+        env:
+          CODECOV_TOKEN: ${{ secrets.CODECOV_TOKEN }}
         run: |
           modal run cicd.multigpu

.github/workflows/tests.yml

@@ -28,6 +28,9 @@ concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}
   cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
 
+permissions:
+  contents: read
+
 env:
   TRANSFORMERS_IS_CI: "yes"
 
@@ -303,9 +306,10 @@ jobs:
           echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
           echo "MODAL_IMAGE_BUILDER_VERSION=2024.10" >> $GITHUB_ENV
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
-          echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
           echo "E2E_DOCKERFILE=${{ matrix.dockerfile || 'Dockerfile.jinja'}}" >> $GITHUB_ENV
       - name: Run tests job on Modal
+        env:
+          CODECOV_TOKEN: ${{ secrets.CODECOV_TOKEN }}
         run: |
           modal run cicd.e2e_tests
 
@@ -371,9 +375,10 @@ jobs:
           echo "MODAL_IMAGE_BUILDER_VERSION=2024.10" >> $GITHUB_ENV
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
           echo "GPU_TYPE=${{ matrix.gpu_type || 'L40S'}}" >> $GITHUB_ENV
-          echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
           echo "E2E_DOCKERFILE=${{ matrix.dockerfile || 'Dockerfile.jinja'}}" >> $GITHUB_ENV
       - name: Run tests job on Modal
+        env:
+          CODECOV_TOKEN: ${{ secrets.CODECOV_TOKEN }}
         run: |
           modal run cicd.e2e_tests
 
@@ -413,7 +418,6 @@ jobs:
           echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
           echo "MODAL_IMAGE_BUILDER_VERSION=2024.10" >> $GITHUB_ENV
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
-          echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
       - name: Run tests job on Modal
         run: |
           modal run cicd.cleanup

README.md

@@ -30,7 +30,7 @@
 ## 🎉 Latest Updates
 
 - 2026/03:
-  - New model support has been added in Axolotl for [Qwen3.5, Qwen3.5 MoE](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/qwen3.5), [GLM-4.7-Flash](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/glm47-flash), [GLM-4.6V](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/glm46v), and [GLM-4.5-Air](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/glm45).
+  - New model support has been added in Axolotl for [Mistral Small 4](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/mistral4), [Qwen3.5, Qwen3.5 MoE](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/qwen3.5), [GLM-4.7-Flash](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/glm47-flash), [GLM-4.6V](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/glm46v), and [GLM-4.5-Air](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/glm45).
   - [MoE expert quantization](https://docs.axolotl.ai/docs/expert_quantization.html) support (via `quantize_moe_experts: true`) greatly reduces VRAM when training MoE models (FSDP2 compat).
 - 2026/02:
   - [ScatterMoE LoRA](https://github.com/axolotl-ai-cloud/axolotl/pull/3410) support. LoRA fine-tuning directly on MoE expert weights using custom Triton kernels.
@@ -75,7 +75,7 @@ Features:
 - **Multimodal Training**: Fine-tune vision-language models (VLMs) including LLaMA-Vision, Qwen2-VL, Pixtral, LLaVA, SmolVLM2, GLM-4.6V, InternVL 3.5, Gemma 3n, and audio models like Voxtral with image, video, and audio support.
 - **Training Methods**: Full fine-tuning, LoRA, QLoRA, GPTQ, QAT, Preference Tuning (DPO, IPO, KTO, ORPO), RL (GRPO, GDPO), and Reward Modelling (RM) / Process Reward Modelling (PRM).
 - **Easy Configuration**: Re-use a single YAML configuration file across the full fine-tuning pipeline: dataset preprocessing, training, evaluation, quantization, and inference.
-- **Performance Optimizations**: [Multipacking](https://docs.axolotl.ai/docs/multipack.html), [Flash Attention](https://github.com/Dao-AILab/flash-attention), [Xformers](https://github.com/facebookresearch/xformers), [Flex Attention](https://pytorch.org/blog/flexattention/), [SageAttention](https://github.com/thu-ml/SageAttention), [Liger Kernel](https://github.com/linkedin/Liger-Kernel), [Cut Cross Entropy](https://github.com/apple/ml-cross-entropy/tree/main), [ScatterMoE](https://docs.axolotl.ai/docs/custom_integrations.html#kernels-integration), [Sequence Parallelism (SP)](https://docs.axolotl.ai/docs/sequence_parallelism.html), [LoRA optimizations](https://docs.axolotl.ai/docs/lora_optims.html), [Multi-GPU training (FSDP1, FSDP2, DeepSpeed)](https://docs.axolotl.ai/docs/multi-gpu.html), [Multi-node training (Torchrun, Ray)](https://docs.axolotl.ai/docs/multi-node.html), and many more!
+- **Performance Optimizations**: [Multipacking](https://docs.axolotl.ai/docs/multipack.html), [Flash Attention 2/3/4](https://docs.axolotl.ai/docs/attention.html#flash-attention), [Xformers](https://docs.axolotl.ai/docs/attention.html#xformers), [Flex Attention](https://docs.axolotl.ai/docs/attention.html#flex-attention), [SageAttention](https://docs.axolotl.ai/docs/attention.html#sageattention), [Liger Kernel](https://docs.axolotl.ai/docs/custom_integrations.html#liger-kernels), [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy), [ScatterMoE](https://docs.axolotl.ai/docs/custom_integrations.html#kernels-integration), [Sequence Parallelism (SP)](https://docs.axolotl.ai/docs/sequence_parallelism.html), [LoRA optimizations](https://docs.axolotl.ai/docs/lora_optims.html), [Multi-GPU training (FSDP1, FSDP2, DeepSpeed)](https://docs.axolotl.ai/docs/multi-gpu.html), [Multi-node training (Torchrun, Ray)](https://docs.axolotl.ai/docs/multi-node.html), and many more!
 - **Flexible Dataset Handling**: Load from local, HuggingFace, and cloud (S3, Azure, GCP, OCI) datasets.
 - **Cloud Ready**: We ship [Docker images](https://hub.docker.com/u/axolotlai) and also [PyPI packages](https://pypi.org/project/axolotl/) for use on cloud platforms and local hardware.
 

_quarto.yml

@@ -128,11 +128,9 @@ quartodoc:
         - monkeypatch.mistral_attn_hijack_flash
         - monkeypatch.multipack
         - monkeypatch.relora
-        - monkeypatch.llama_expand_mask
         - monkeypatch.lora_kernels
         - monkeypatch.utils
         - monkeypatch.btlm_attn_hijack_flash
-        - monkeypatch.llama_patch_multipack
         - monkeypatch.stablelm_attn_hijack_flash
         - monkeypatch.trainer_fsdp_optim
         - monkeypatch.transformers_fa_utils

benchmarks/bench_entropy.py

@@ -0,0 +1,208 @@
+"""Benchmark for entropy_from_logits Triton kernel vs original chunked implementation.
+
+Usage: CUDA_VISIBLE_DEVICES=0 python benchmarks/bench_entropy.py
+"""
+
+import gc
+import statistics
+
+import torch
+import torch.nn.functional as F
+
+from axolotl.monkeypatch.trainer.utils import entropy_from_logits
+
+V = 151936  # Qwen vocab
+WARMUP = 5
+BENCH_ITERS = 20
+MEM_ITERS = 10
+
+
+def entropy_from_logits_original(logits: torch.Tensor, chunk_size: int = 128):
+    """Original chunked implementation (reference)."""
+    original_shape = logits.shape[:-1]
+    num_classes = logits.shape[-1]
+    flat_logits = logits.reshape(-1, num_classes)
+    entropies = []
+    for chunk in flat_logits.split(chunk_size, dim=0):
+        logps = F.log_softmax(chunk, dim=-1)
+        chunk_entropy = -(torch.exp(logps) * logps).sum(-1)
+        entropies.append(chunk_entropy)
+    return torch.cat(entropies, dim=0).reshape(original_shape)
+
+
+def _clean_gpu():
+    gc.collect()
+    torch.cuda.empty_cache()
+    torch.cuda.reset_peak_memory_stats()
+    torch.cuda.reset_accumulated_memory_stats()
+    torch.cuda.synchronize()
+
+
+def profile_time(fn, logits, n_iters=BENCH_ITERS):
+    for _ in range(WARMUP):
+        out = fn(logits, chunk_size=128)
+        del out
+    torch.cuda.synchronize()
+
+    times = []
+    for _ in range(n_iters):
+        s = torch.cuda.Event(enable_timing=True)
+        e = torch.cuda.Event(enable_timing=True)
+        s.record()
+        out = fn(logits, chunk_size=128)
+        e.record()
+        torch.cuda.synchronize()
+        times.append(s.elapsed_time(e))
+        del out
+    return times
+
+
+def profile_memory(fn, logits, n_iters=MEM_ITERS):
+    for _ in range(WARMUP):
+        out = fn(logits, chunk_size=128)
+        del out
+    torch.cuda.synchronize()
+
+    peaks = []
+    for _ in range(n_iters):
+        _clean_gpu()
+        base = torch.cuda.max_memory_allocated()
+        out = fn(logits, chunk_size=128)
+        torch.cuda.synchronize()
+        peaks.append(torch.cuda.max_memory_allocated() - base)
+        del out
+    return [p / 1e6 for p in peaks]
+
+
+def fmt(values, unit=""):
+    mean = statistics.mean(values)
+    std = statistics.stdev(values) if len(values) > 1 else 0.0
+    return f"{mean:8.2f} ± {std:5.2f} {unit}  [min={min(values):.2f}, max={max(values):.2f}]"
+
+
+def benchmark_contiguous():
+    print("=" * 60)
+    print(
+        f"CONTIGUOUS BENCHMARK  (warmup={WARMUP}, time={BENCH_ITERS}, mem={MEM_ITERS})"
+    )
+    print("=" * 60)
+
+    configs = [
+        (1, 2048),
+        (1, 8192),
+        (1, 16384),
+        (4, 4096),
+        (8, 2048),
+        (16, 2048),
+        (16, 4096),
+    ]
+
+    for B, L in configs:
+        mem_gb = B * L * V * 2 / 1e9
+        if mem_gb > 28:
+            print(f"\n  skip B={B}, L={L} ({mem_gb:.1f} GB)")
+            continue
+
+        N = B * L
+        print(f"\n{'─' * 60}")
+        print(f"B={B:2d}, L={L:5d}  ({N:6d} rows, logits {mem_gb:.2f} GB)")
+        print(f"{'─' * 60}")
+
+        torch.manual_seed(42)
+        logits = torch.randn(B, L, V, device="cuda", dtype=torch.bfloat16)
+
+        t_orig = profile_time(entropy_from_logits_original, logits)
+        t_triton = profile_time(entropy_from_logits, logits)
+        orig_mean = statistics.mean(t_orig)
+        triton_mean = statistics.mean(t_triton)
+
+        print("  TIME (ms):")
+        print(f"    original: {fmt(t_orig, 'ms')}")
+        print(f"    triton:   {fmt(t_triton, 'ms')}")
+        print(f"    speedup:  {orig_mean / triton_mean:.2f}x")
+
+        m_orig = profile_memory(entropy_from_logits_original, logits)
+        m_triton = profile_memory(entropy_from_logits, logits)
+        orig_peak = statistics.mean(m_orig)
+        triton_peak = statistics.mean(m_triton)
+
+        print("  MEMORY (peak overhead):")
+        print(f"    original: {fmt(m_orig, 'MB')}")
+        print(f"    triton:   {fmt(m_triton, 'MB')}")
+        print(f"    saved:    {orig_peak - triton_peak:.1f} MB")
+
+        del logits
+        _clean_gpu()
+
+
+def benchmark_noncontiguous():
+    print("\n" + "=" * 60)
+    print(
+        f"NON-CONTIGUOUS BENCHMARK  (warmup={WARMUP}, time={BENCH_ITERS}, mem={MEM_ITERS})"
+    )
+    print("=" * 60)
+
+    configs = [
+        (4, 2048, "transpose"),
+        (4, 8192, "transpose"),
+        (8, 2048, "transpose"),
+        (4, 4096, "slice_batch"),
+    ]
+
+    for B, L, method in configs:
+        torch.manual_seed(42)
+
+        if method == "transpose":
+            raw = torch.randn(L, B, V, device="cuda", dtype=torch.bfloat16)
+            logits_nc = raw.transpose(0, 1)
+            raw_gb = L * B * V * 2 / 1e9
+        elif method == "slice_batch":
+            raw = torch.randn(B * 2, L, V, device="cuda", dtype=torch.bfloat16)
+            logits_nc = raw[::2]
+            raw_gb = B * 2 * L * V * 2 / 1e9
+        else:
+            continue
+
+        if raw_gb > 28:
+            print(f"\n  skip B={B}, L={L}, {method} ({raw_gb:.1f} GB)")
+            del raw, logits_nc
+            torch.cuda.empty_cache()
+            continue
+
+        N = B * L
+        print(f"\n{'─' * 60}")
+        print(f"B={B}, L={L}  {method}  ({N} rows, raw {raw_gb:.2f} GB)")
+        print(f"{'─' * 60}")
+
+        def original_with_copy(logits, chunk_size=128):
+            return entropy_from_logits_original(
+                logits.contiguous(), chunk_size=chunk_size
+            )
+
+        t_orig = profile_time(original_with_copy, logits_nc)
+        t_triton = profile_time(entropy_from_logits, logits_nc)
+        orig_mean = statistics.mean(t_orig)
+        triton_mean = statistics.mean(t_triton)
+
+        print("  TIME (ms):")
+        print(f"    orig+copy:     {fmt(t_orig, 'ms')}")
+        print(f"    triton-strided:{fmt(t_triton, 'ms')}")
+        print(f"    speedup:       {orig_mean / triton_mean:.2f}x")
+
+        m_orig = profile_memory(original_with_copy, logits_nc)
+        m_triton = profile_memory(entropy_from_logits, logits_nc)
+        orig_peak = statistics.mean(m_orig)
+        triton_peak = statistics.mean(m_triton)
+
+        print("  MEMORY (peak overhead):")
+        print(f"    orig+copy:     {fmt(m_orig, 'MB')}")
+        print(f"    triton-strided:{fmt(m_triton, 'MB')}")
+        print(f"    saved:         {orig_peak - triton_peak:.1f} MB")
+
+        del raw, logits_nc
+        _clean_gpu()
+
+
+if __name__ == "__main__":
+    benchmark_contiguous()
+    benchmark_noncontiguous()

benchmarks/bench_scattermoe_lora.py

@@ -0,0 +1,284 @@
+"""Benchmark for ScatterMoE LoRA Triton kernels.
+
+Measures forward, backward dX, and backward dA/dB kernels at common MoE
+model shapes. Reports per-kernel timings, LoRA overhead vs base scatter2scatter,
+and full fwd+bwd autograd throughput.
+
+Usage:
+  CUDA_VISIBLE_DEVICES=0 python benchmarks/bench_scattermoe_lora.py
+  CUDA_VISIBLE_DEVICES=0 python benchmarks/bench_scattermoe_lora.py --ranks 16 64
+  CUDA_VISIBLE_DEVICES=0 python benchmarks/bench_scattermoe_lora.py --models Qwen/Qwen3.5-35B-A3B
+"""
+
+import argparse
+import gc
+import time
+from functools import partial
+
+import torch
+
+from axolotl.integrations.kernels.libs.scattermoe_lora.kernels import (
+    lora_ops,
+    ops as base_ops,
+)
+from axolotl.integrations.kernels.libs.scattermoe_lora.parallel_experts import (
+    flatten_sort_count,
+)
+from axolotl.integrations.kernels.libs.scattermoe_lora.parallel_linear_lora import (
+    ScatterMoELoRA,
+)
+
+DEVICE = "cuda"
+DTYPE = torch.bfloat16
+WARMUP = 5
+ITERS = 20
+
+# ─── Model configs ──────────────────────────────────────────────────────────
+
+BUILTIN_CONFIGS = {
+    "Qwen3.5-35B-A3B": (256, 2048, 512, 8),  # E, H, I, k
+    "Qwen3-30B-A3B": (128, 2048, 768, 8),
+    "OLMoE-1B-7B": (64, 2048, 1024, 8),
+    "Mixtral-8x7B": (8, 4096, 14336, 2),
+}
+
+
+def _resolve_config(spec):
+    """Resolve a model spec to (E, H, I, k). Accepts builtin names or HF IDs."""
+    key = spec.lower().replace("/", "-")
+    for name, cfg in BUILTIN_CONFIGS.items():
+        if key in name.lower() or name.lower() in key:
+            return name, cfg
+
+    from transformers import AutoConfig
+
+    hf_cfg = AutoConfig.from_pretrained(spec, trust_remote_code=True)
+    if callable(getattr(hf_cfg, "get_text_config", None)):
+        tc = hf_cfg.get_text_config()
+        if hasattr(tc, "model_type") and tc.model_type != hf_cfg.model_type:
+            hf_cfg = tc
+    hidden = hf_cfg.hidden_size
+    inter = getattr(hf_cfg, "moe_intermediate_size", None) or hf_cfg.intermediate_size
+    experts = (
+        getattr(hf_cfg, "num_experts", None)
+        or getattr(hf_cfg, "num_local_experts", None)
+        or getattr(hf_cfg, "n_routed_experts", None)
+    )
+    top_k = (
+        getattr(hf_cfg, "num_experts_per_tok", None)
+        or getattr(hf_cfg, "num_experts_per_token", None)
+        or 2
+    )
+    name = spec.split("/")[-1]
+    return name, (experts, hidden, inter, top_k)
+
+
+# ─── Benchmark helpers ──────────────────────────────────────────────────────
+
+
+def _clean():
+    gc.collect()
+    torch.cuda.empty_cache()
+    torch.cuda.synchronize()
+
+
+def _bench(fn, warmup=WARMUP, iters=ITERS):
+    for _ in range(warmup):
+        fn()
+    torch.cuda.synchronize()
+    times = []
+    for _ in range(iters):
+        torch.cuda.synchronize()
+        t0 = time.perf_counter()
+        fn()
+        torch.cuda.synchronize()
+        times.append((time.perf_counter() - t0) * 1000)
+    times.sort()
+    return times[len(times) // 2]
+
+
+def _setup(num_experts, K, N, T, top_k, R):
+    torch.manual_seed(42)
+    x = torch.randn(T, K, device=DEVICE, dtype=DTYPE)
+    W = torch.randn(num_experts, K, N, device=DEVICE, dtype=DTYPE) * 0.02
+    lora_A = torch.randn(R * num_experts, K, device=DEVICE, dtype=DTYPE) * 0.01
+    lora_B = torch.randn(N, R * num_experts, device=DEVICE, dtype=DTYPE) * 0.01
+    logits = torch.randn(T, num_experts, device=DEVICE)
+    _, top_idx = torch.topk(torch.softmax(logits, dim=-1), top_k, dim=-1)
+    sei, ssi, eo = flatten_sort_count(top_idx, num_experts)
+    gx = base_ops.group(x, ssi, fan_out=top_k)
+    dy = torch.randn(gx.size(0), N, device=DEVICE, dtype=DTYPE)
+    return x, W, lora_A, lora_B, sei, ssi, eo, gx, dy
+
+
+# ─── Kernel wrappers (avoid B023 loop-variable capture) ──────────────────────
+
+
+def _call_fwd(x, W, sei, ssi, top_k, lA, lB):
+    return lora_ops.scatter2scatter_lora(
+        X=x,
+        W=W,
+        sorted_expert_idxs=sei,
+        sorted_scattered_idxs=ssi,
+        k=top_k,
+        lora_A=lA,
+        lora_B=lB,
+        scaling=2.0,
+    )
+
+
+def _call_base(x, W, sei, ssi, top_k):
+    return base_ops.scatter2scatter(
+        X=x,
+        W=W,
+        sorted_expert_idxs=sei,
+        sorted_scattered_idxs=ssi,
+        k=top_k,
+    )
+
+
+def _call_dx(dy, W, sei, ssi, lA, lB):
+    return lora_ops.scatter2scatter_lora_dX(
+        DY=dy,
+        W=W,
+        sorted_expert_idxs=sei,
+        sorted_scattered_idxs=ssi,
+        k=1,
+        lora_A=lA,
+        lora_B=lB,
+        scaling=2.0,
+        dy_grouped=True,
+        dx_grouped=False,
+    )
+
+
+def _call_bwd(dy, gx, lA, lB, eo, num_experts):
+    return lora_ops.group_bwd_lora(
+        DY=dy,
+        X=gx,
+        lora_A=lA,
+        lora_B=lB,
+        expert_offsets=eo,
+        E=num_experts,
+        scaling=2.0,
+    )
+
+
+# ─── Main ────────────────────────────────────────────────────────────────────
+
+
+def main():
+    parser = argparse.ArgumentParser(description="ScatterMoE LoRA kernel benchmark")
+    parser.add_argument(
+        "--models",
+        "-m",
+        nargs="+",
+        help="Model names or HF IDs (default: all builtins)",
+    )
+    parser.add_argument("--ranks", "-r", nargs="+", type=int, default=[16, 32, 64])
+    parser.add_argument("--seq-len", "-T", type=int, default=2048)
+    args = parser.parse_args()
+
+    T = args.seq_len
+    print(f"GPU: {torch.cuda.get_device_name()}")
+    print(f"T={T}, ranks={args.ranks}\n")
+
+    if args.models:
+        configs = [_resolve_config(m) for m in args.models]
+    else:
+        configs = list(BUILTIN_CONFIGS.items())
+
+    for model_name, (num_experts, hidden, inter, top_k) in configs:
+        print(f"{'=' * 70}")
+        print(f"  {model_name}: E={num_experts}, H={hidden}, I={inter}, k={top_k}")
+        print(f"{'=' * 70}")
+
+        for R in args.ranks:
+            for proj, K, N in [("gate_up", hidden, 2 * inter), ("down", inter, hidden)]:
+                _clean()
+                x, W, lA, lB, sei, ssi, eo, gx, dy = _setup(
+                    num_experts, K, N, T, top_k, R
+                )
+
+                # Forward with LoRA (auto-dispatched: fused or split)
+                dispatch = (
+                    "split"
+                    if (
+                        num_experts <= lora_ops._SPLIT_LORA_FWD_MAX_EXPERTS
+                        and K * N >= lora_ops._SPLIT_LORA_FWD_THRESHOLD
+                    )
+                    else "fused"
+                )
+                t_fwd = _bench(partial(_call_fwd, x, W, sei, ssi, top_k, lA, lB))
+                t_base = _bench(partial(_call_base, x, W, sei, ssi, top_k))
+                t_dx = _bench(partial(_call_dx, dy, W, sei, ssi, lA, lB))
+                t_bwd = _bench(partial(_call_bwd, dy, gx, lA, lB, eo, num_experts))
+
+                total = t_fwd + t_dx + t_bwd
+                overhead = t_fwd / t_base - 1 if t_base > 0 else 0
+
+                print(
+                    f"  R={R:>2} {proj:<8}  "
+                    f"fwd={t_fwd:>6.2f}ms [{dispatch}]  "
+                    f"base={t_base:>6.2f}ms "
+                    f"(+{overhead * 100:.0f}%)  "
+                    f"dx={t_dx:>6.2f}ms  bwd={t_bwd:>6.2f}ms  "
+                    f"total={total:>6.2f}ms"
+                )
+
+                # Full autograd fwd+bwd with memory measurement
+                x_ag = x.clone().requires_grad_(True)
+                lA_ag = lA.clone().requires_grad_(True)
+                lB_ag = lB.clone().requires_grad_(True)
+
+                def _run_autograd(
+                    _x=x_ag,
+                    _W=W,
+                    _k=top_k,
+                    _sei=sei,
+                    _ssi=ssi,
+                    _eo=eo,
+                    _lA=lA_ag,
+                    _lB=lB_ag,
+                ):
+                    out = ScatterMoELoRA.apply(
+                        _x,
+                        _W,
+                        _k,
+                        _sei,
+                        _ssi,
+                        _eo,
+                        _lA,
+                        _lB,
+                        2.0,
+                        None,
+                        None,
+                        False,
+                        False,
+                        True,
+                        False,
+                    )
+                    out.sum().backward()
+                    _x.grad = None
+                    _lA.grad = None
+                    _lB.grad = None
+
+                t_full = _bench(_run_autograd)
+
+                _clean()
+                torch.cuda.reset_peak_memory_stats()
+                mem_before = torch.cuda.memory_allocated()
+                _run_autograd()
+                torch.cuda.synchronize()
+                mem_peak = torch.cuda.max_memory_allocated() - mem_before
+
+                print(
+                    f"         full_fwd_bwd={t_full:>6.2f}ms  "
+                    f"peak_delta={mem_peak / 1e6:>6.1f}MB"
+                )
+
+        print()
+
+
+if __name__ == "__main__":
+    main()

benchmarks/bench_selective_logsoftmax.py

@@ -0,0 +1,191 @@
+"""Benchmark for selective_log_softmax Triton kernel vs original implementation.
+
+Usage: CUDA_VISIBLE_DEVICES=0 python benchmarks/bench_selective_logsoftmax.py
+"""
+
+import gc
+import statistics
+
+import torch
+
+from axolotl.monkeypatch.trainer.utils import (
+    selective_log_softmax,
+    selective_log_softmax_original,
+)
+
+V = 151936  # Qwen vocab
+WARMUP = 5
+BENCH_ITERS = 20
+MEM_ITERS = 10
+
+
+def _clean_gpu():
+    gc.collect()
+    torch.cuda.empty_cache()
+    torch.cuda.reset_peak_memory_stats()
+    torch.cuda.reset_accumulated_memory_stats()
+    torch.cuda.synchronize()
+
+
+def profile_time(fn, args, n_iters=BENCH_ITERS):
+    for _ in range(WARMUP):
+        fn(*args)
+    torch.cuda.synchronize()
+
+    times = []
+    for _ in range(n_iters):
+        s = torch.cuda.Event(enable_timing=True)
+        e = torch.cuda.Event(enable_timing=True)
+        s.record()
+        fn(*args)
+        e.record()
+        torch.cuda.synchronize()
+        times.append(s.elapsed_time(e))
+    return times
+
+
+def profile_memory(fn, args, n_iters=MEM_ITERS):
+    for _ in range(WARMUP):
+        out = fn(*args)
+        del out
+    torch.cuda.synchronize()
+
+    peaks = []
+    for _ in range(n_iters):
+        _clean_gpu()
+        base = torch.cuda.max_memory_allocated()
+        out = fn(*args)
+        torch.cuda.synchronize()
+        peaks.append(torch.cuda.max_memory_allocated() - base)
+        del out
+    return [p / 1e6 for p in peaks]
+
+
+def fmt(values, unit=""):
+    mean = statistics.mean(values)
+    std = statistics.stdev(values) if len(values) > 1 else 0.0
+    return f"{mean:8.2f} ± {std:5.2f} {unit}  [min={min(values):.2f}, max={max(values):.2f}]"
+
+
+def benchmark_forward():
+    print("=" * 60)
+    print(f"FORWARD BENCHMARK  (warmup={WARMUP}, time={BENCH_ITERS}, mem={MEM_ITERS})")
+    print("=" * 60)
+
+    configs = [
+        (1, 2048),
+        (1, 8192),
+        (4, 4096),
+        (8, 2048),
+        (16, 2048),
+        (16, 4096),
+    ]
+
+    for B, L in configs:
+        mem_gb = B * L * V * 2 / 1e9
+        if mem_gb > 28:
+            print(f"\n  skip B={B}, L={L} ({mem_gb:.1f} GB)")
+            continue
+
+        N = B * L
+        print(f"\n{'─' * 60}")
+        print(f"B={B:2d}, L={L:5d}  ({N:6d} rows, logits {mem_gb:.2f} GB)")
+        print(f"{'─' * 60}")
+
+        torch.manual_seed(42)
+        logits = torch.randn(B, L, V, device="cuda", dtype=torch.bfloat16)
+        index = torch.randint(0, V, (B, L), device="cuda")
+
+        t_orig = profile_time(selective_log_softmax_original, (logits, index))
+        t_triton = profile_time(selective_log_softmax, (logits, index))
+        orig_mean = statistics.mean(t_orig)
+        triton_mean = statistics.mean(t_triton)
+
+        print("  TIME (ms):")
+        print(f"    original: {fmt(t_orig, 'ms')}")
+        print(f"    triton:   {fmt(t_triton, 'ms')}")
+        print(f"    speedup:  {orig_mean / triton_mean:.2f}x")
+
+        m_orig = profile_memory(selective_log_softmax_original, (logits, index))
+        m_triton = profile_memory(selective_log_softmax, (logits, index))
+        orig_peak = statistics.mean(m_orig)
+        triton_peak = statistics.mean(m_triton)
+
+        print("  MEMORY (peak overhead):")
+        print(f"    original: {fmt(m_orig, 'MB')}")
+        print(f"    triton:   {fmt(m_triton, 'MB')}")
+        print(f"    saved:    {orig_peak - triton_peak:.1f} MB")
+
+        del logits, index
+        _clean_gpu()
+
+
+def benchmark_backward():
+    print("\n" + "=" * 60)
+    print(f"FWD+BWD BENCHMARK  (warmup={WARMUP}, time={BENCH_ITERS}, mem={MEM_ITERS})")
+    print("=" * 60)
+
+    configs = [
+        (1, 2048),
+        (1, 8192),
+        (4, 4096),
+        (8, 2048),
+        (16, 2048),
+        (16, 4096),
+    ]
+
+    def fwd_bwd_original(logits, index):
+        logits.grad = None
+        out = selective_log_softmax_original(logits, index)
+        out.sum().backward()
+
+    def fwd_bwd_triton(logits, index):
+        logits.grad = None
+        out = selective_log_softmax(logits, index)
+        out.sum().backward()
+
+    for B, L in configs:
+        mem_gb = B * L * V * 2 / 1e9
+        if mem_gb > 20:
+            print(f"\n  skip B={B}, L={L} ({mem_gb:.1f} GB, need room for grads)")
+            continue
+
+        N = B * L
+        print(f"\n{'─' * 60}")
+        print(f"B={B:2d}, L={L:5d}  ({N:6d} rows, logits {mem_gb:.2f} GB)")
+        print(f"{'─' * 60}")
+
+        torch.manual_seed(42)
+        logits_orig = torch.randn(
+            B, L, V, device="cuda", dtype=torch.bfloat16, requires_grad=True
+        )
+        logits_tri = logits_orig.detach().clone().requires_grad_(True)
+        index = torch.randint(0, V, (B, L), device="cuda")
+
+        t_orig = profile_time(fwd_bwd_original, (logits_orig, index))
+        t_triton = profile_time(fwd_bwd_triton, (logits_tri, index))
+        orig_mean = statistics.mean(t_orig)
+        triton_mean = statistics.mean(t_triton)
+
+        print("  FWD+BWD TIME (ms):")
+        print(f"    original: {fmt(t_orig, 'ms')}")
+        print(f"    triton:   {fmt(t_triton, 'ms')}")
+        print(f"    speedup:  {orig_mean / triton_mean:.2f}x")
+
+        m_orig = profile_memory(fwd_bwd_original, (logits_orig, index))
+        m_triton = profile_memory(fwd_bwd_triton, (logits_tri, index))
+        orig_peak = statistics.mean(m_orig)
+        triton_peak = statistics.mean(m_triton)
+
+        print("  FWD+BWD MEMORY (peak overhead):")
+        print(f"    original: {fmt(m_orig, 'MB')}")
+        print(f"    triton:   {fmt(m_triton, 'MB')}")
+        print(f"    saved:    {orig_peak - triton_peak:.1f} MB")
+
+        del logits_orig, logits_tri, index
+        _clean_gpu()
+
+
+if __name__ == "__main__":
+    benchmark_forward()
+    benchmark_backward()

cicd/Dockerfile-uv.jinja

@@ -11,7 +11,7 @@ ENV NIGHTLY_BUILD="{{ NIGHTLY_BUILD }}"
 ENV HF_HOME="{{ HF_HOME }}"
 
 RUN apt-get update && \
-    apt-get install -y --allow-change-held-packages vim curl nano libnccl2 libnccl-dev ibverbs-providers ibverbs-utils infiniband-diags librdmacm-dev librdmacm1 rdmacm-utils slurm-wlm
+    apt-get install -y --allow-change-held-packages vim curl nano zstd libnccl2 libnccl-dev ibverbs-providers ibverbs-utils infiniband-diags librdmacm-dev librdmacm1 rdmacm-utils slurm-wlm
 
 WORKDIR /workspace
 

cicd/Dockerfile.jinja

@@ -12,7 +12,7 @@ ENV HF_HOME="{{ HF_HOME }}"
 ENV AXOLOTL_DATASET_NUM_PROC="8"
 
 RUN apt-get update && \
-    apt-get install -y --allow-change-held-packages vim curl nano libnccl2 libnccl-dev ibverbs-providers ibverbs-utils infiniband-diags librdmacm-dev librdmacm1 rdmacm-utils slurm-wlm
+    apt-get install -y --allow-change-held-packages vim curl nano zstd libnccl2 libnccl-dev ibverbs-providers ibverbs-utils infiniband-diags librdmacm-dev librdmacm1 rdmacm-utils slurm-wlm
 
 WORKDIR /workspace
 

cicd/cicd.sh

@@ -3,11 +3,13 @@ set -e
 
 python -c "import torch; assert '$PYTORCH_VERSION' in torch.__version__"
 
-# curl -L https://axolotl-ci.b-cdn.net/hf-cache.tar.zst | tar -xpf - -C "${HF_HOME}/hub/"  --use-compress-program unzstd --strip-components=1
-hf download "NousResearch/Meta-Llama-3-8B"
-hf download "NousResearch/Meta-Llama-3-8B-Instruct"
-hf download "microsoft/Phi-4-reasoning"
-hf download "microsoft/Phi-3.5-mini-instruct"
+set -o pipefail
+curl --silent --show-error --fail --retry 3 --retry-delay 5 -L https://axolotl-ci.b-cdn.net/hf-cache.tar.zst | tar -xpf - -C "${HF_HOME}/hub/" --use-compress-program unzstd --strip-components=1
+# hf download "NousResearch/Meta-Llama-3-8B"
+# hf download "NousResearch/Meta-Llama-3-8B-Instruct"
+# hf download "microsoft/Phi-4-reasoning"
+# hf download "microsoft/Phi-3.5-mini-instruct"
+# hf download "microsoft/Phi-3-medium-128k-instruct"
 
 # Run unit tests with initial coverage report
 pytest -v --durations=10 -n8 \

cicd/single_gpu.py

@@ -68,10 +68,6 @@ def run_cmd(cmd: str, run_folder: str):
     sp_env["AXOLOTL_DATASET_NUM_PROC"] = "8"
 
     # Propagate errors from subprocess.
-    try:
-        exit_code = subprocess.call(cmd.split(), cwd=run_folder, env=sp_env)  # nosec
-        if exit_code:
-            print(f"Command '{cmd}' failed with exit code {exit_code}")
-            return exit_code
-    except Exception as e:  # pylint: disable=broad-except
-        print(f"Command '{cmd}' failed with exception {e}")
+    exit_code = subprocess.call(cmd.split(), cwd=run_folder, env=sp_env)  # nosec
+    if exit_code:
+        raise RuntimeError(f"Command '{cmd}' failed with exit code {exit_code}")

codecov.yml

@@ -37,6 +37,7 @@ coverage:
         only_pulls: false
         flags: null
         paths: null
+        informational: true
 
 parsers:
   gcov:

docs/attention.qmd

@@ -13,9 +13,10 @@ sdp_attention: true
 
 For more details: [PyTorch docs](https://docs.pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html)
 
-## Flash Attention 2
+## Flash Attention
 
-Uses efficient kernels to compute attention.
+Axolotl supports Flash Attention 2, 3, and 4. The best available version is used automatically
+based on your installed packages and GPU.
 
 ```yaml
 flash_attention: true
@@ -23,11 +24,9 @@ flash_attention: true
 
 For more details: [Flash Attention](https://github.com/Dao-AILab/flash-attention/)
 
-### Nvidia
+### Flash Attention 2
 
-Requirements: Ampere, Ada, or Hopper GPUs
-
-Note: For Turing GPUs or lower, please use other attention methods.
+Requirements: Ampere, Ada, or Hopper GPUs (Turing or lower not supported)
 
 ```bash
 pip install flash-attn --no-build-isolation
@@ -35,11 +34,12 @@ pip install flash-attn --no-build-isolation
 
 ::: {.callout-tip}
 
-If you get `undefined symbol` while training, ensure you installed PyTorch prior to Axolotl. Alternatively, try reinstall or downgrade a version.
+If you get `undefined symbol` while training, ensure you installed PyTorch prior to Axolotl.
+Alternatively, try reinstall or downgrade a version.
 
 :::
 
-#### Flash Attention 3
+### Flash Attention 3
 
 Requirements: Hopper only and CUDA 12.8 (recommended)
 
@@ -50,6 +50,44 @@ cd flash-attention/hopper
 python setup.py install
 ```
 
+### Flash Attention 4
+
+Requirements: Hopper or Blackwell GPUs
+
+```bash
+pip install flash-attn-4
+```
+
+Or from source:
+
+```bash
+git clone https://github.com/Dao-AILab/flash-attention.git
+cd flash-attention/flash_attn/cute
+
+pip install -e .
+
+# FA2's flash_attn package includes a cute/ stub that shadows FA4.
+# Remove it so Python can find the real FA4 module:
+rm -r $(python -c "import flash_attn; print(flash_attn.__path__[0])")/cute
+```
+
+::: {.callout-note}
+
+**Hopper (SM90) users**: The backward kernel is not yet included in the pip package. To use FA4
+for training on Hopper, install from source using the instructions above.
+
+:::
+
+::: {.callout-warning}
+
+FA4 only supports head dimensions up to 128 (`d ≤ 128`). The DeepSeek shape `(192, 128)` is
+also supported but only on Blackwell. Axolotl automatically detects incompatible head dimensions
+and falls back to FA2/3.
+
+:::
+
+For more details: [flash-attention/flash_attn/cute](https://github.com/Dao-AILab/flash-attention/tree/main/flash_attn/cute)
+
 ### AMD
 
 Requirements: ROCm 6.0 and above.

docs/gradient_checkpointing.qmd

@@ -1,5 +1,5 @@
 ---
-title: Gradient Checkpointing and Activation Offloading
+title: Gradient Checkpointing, Activation Offloading, and Layer Offloading
 ---
 
 Gradient checkpointing and activation offloading are techniques used to optimize the performance of deep learning
@@ -27,3 +27,33 @@ The `activation_offloading: legacy` naively offloads activations to CPU and with
 
 For resource constrained environments with limited CPU memory, `activation_offloading: disk` offloads
 activations to disk instead of CPU RAM so that much larger context lengths can be trained with minimal memory.
+
+### Enabling Layer Offloading
+
+```yaml
+layer_offloading: true
+```
+
+Layer offloading reduces GPU memory usage by moving frozen (non-trainable) decoder layer parameters to CPU
+and streaming them back to GPU one layer at a time during the forward and backward passes. This is
+particularly useful for LoRA/QLoRA training where most of the model's parameters are frozen — only the
+trainable adapter weights stay on GPU permanently.
+
+During training, forward and backward hooks on each decoder layer handle the transfer automatically:
+
+- **Forward pass:** Before a layer executes, its frozen params are loaded to GPU. The next layer is
+  prefetched asynchronously on a separate CUDA stream for overlap.
+- **Backward pass:** Same pattern in reverse — the current layer's frozen params are loaded and the
+  previous layer is prefetched.
+
+After each layer finishes, its frozen params are offloaded back to CPU pinned memory.
+
+This approach trades some CPU-GPU transfer overhead for significant GPU memory savings — the freed memory
+is roughly equal to the size of all frozen parameters across all decoder layers, minus one layer's worth
+that is kept on GPU at any given time.
+
+**Requirements:**
+
+- CUDA GPU (CPU-only training is not supported for this feature)
+- Works with any HuggingFace model architecture that uses decoder layers (Llama, Mistral, Qwen, etc.)
+- Best combined with LoRA/QLoRA where most parameters are frozen

docs/multimodal.qmd

@@ -13,12 +13,14 @@ format:
 - [Pixtral](#sec-pixtral)
 - [Llava-1.5](#sec-llava-15)
 - [Mistral-Small-3.1](#sec-mistral-small-31)
+- [Mistral-Small-4](#sec-mistral-small-4)
 - [Magistral-Small-2509](#sec-magistral-small-2509)
 - [Voxtral](#sec-voxtral)
 - [Gemma-3](#sec-gemma-3)
 - [Gemma-3n](#sec-gemma-3n)
 - [Qwen2-VL](#sec-qwen2-vl)
 - [Qwen2.5-VL](#sec-qwen25-vl)
+- [Qwen3.5](#sec-qwen3-5)
 - [GLM-4.6V](#sec-glm-4-6v)
 - [SmolVLM2](#sec-smolvlm2)
 - [LFM2-VL](#sec-lfm2-vl)
@@ -108,6 +110,12 @@ Please make sure to install vision lib via `pip install 'mistral-common[opencv]=
 base_model: mistralai/Mistral-Small-3.1-24B-Instruct-2503
 ```
 
+### Mistral-Small-4 {#sec-mistral-small-4}
+
+```yaml
+base_model: mistralai/Mistral-Small-4-119B-2603
+```
+
 ### Magistral-Small-2509 {#sec-magistral-small-2509}
 
 ::: {.callout-tip}
@@ -184,6 +192,14 @@ base_model: Qwen/Qwen3-VL-4B-Instruct
 chat_template: qwen2_vl  # same as qwen2-vl
 ```
 
+### Qwen3.5 {#sec-qwen3-5}
+
+```yaml
+base_model: Qwen/Qwen3.5-9B
+
+chat_template: qwen3_5
+```
+
 ### GLM-4.6V {#sec-glm-4-6v}
 
 Both GLM-4.6V (106B MoE) and GLM-4.6V-Flash (9B) are supported.

docs/optimizations.qmd

@@ -54,6 +54,13 @@ These techniques save VRAM by changing how activations are handled.
 - Activation Offloading: moves activations to CPU RAM or disk, trading I/O overhead for VRAM.
 - Learn more: [Gradient Checkpointing and Offloading Docs](gradient_checkpointing.qmd)
 
+### Layer Offloading
+
+Offloads frozen (non-trainable) decoder layer parameters to CPU and streams them back to GPU one layer at a time during forward/backward passes using CUDA stream prefetching. Especially effective for LoRA/QLoRA where most parameters are frozen.
+
+- **Config:** `layer_offloading: true`
+- **Learn more:** [Layer Offloading Docs](gradient_checkpointing.qmd#enabling-layer-offloading)
+
 ### Cut Cross Entropy (CCE)
 
 Reduces VRAM usage by using an optimized cross-entropy loss calculation.

docs/rlhf.qmd

@@ -721,6 +721,213 @@ trl:
 
 For more information, see [GRPO docs](https://huggingface.co/docs/trl/v0.17.0/en/grpo_trainer#loss-types).
 
+#### Async GRPO
+
+Async GRPO overlaps vLLM generation with training by producing rollouts in a background thread. While the model trains on the current batch, the next batch is already being generated. This can significantly reduce wall-clock time per step.
+
+```yaml
+trl:
+  use_data_producer: true     # Enable data producer protocol
+  use_vllm: true
+  async_prefetch: true         # Generate rollouts in background thread
+  prefetch_depth: 1            # Number of rollouts to prefetch
+  vllm_sync_interval: 2        # Sync weights to vLLM every N steps
+```
+
+::: {.callout-note}
+Because the background thread generates completions with slightly stale model weights, async GRPO uses importance sampling correction to account for the distribution shift. This is controlled by `vllm_importance_sampling_correction: true` (default when async is enabled).
+:::
+
+##### vLLM LoRA Sync
+
+By default, weight sync to vLLM merges the LoRA adapter into the base model and broadcasts all parameters via NCCL. LoRA sync is a faster alternative that saves only the adapter weights to the filesystem and has vLLM load them natively using Punica kernels.
+
+```yaml
+adapter: lora
+lora_r: 32
+lora_alpha: 64
+lora_target_linear: true
+
+trl:
+  vllm_lora_sync: true         # Enable native LoRA sync
+```
+
+When `vllm_lora_sync: true` is set, axolotl automatically selects the LoRA-aware vLLM serve module. Start vLLM as usual:
+
+```bash
+CUDA_VISIBLE_DEVICES=0 axolotl vllm-serve config.yaml
+```
+
+Then start training on a separate GPU:
+
+```bash
+CUDA_VISIBLE_DEVICES=1 axolotl train config.yaml
+```
+
+::: {.callout-tip}
+LoRA sync is especially beneficial with multi-GPU training (FSDP/DeepSpeed), where NCCL merge-sync can cause GPU contention with vLLM generation.
+:::
+
+##### Streaming Partial Batch
+
+Instead of scoring the entire batch at once, streaming mode scores one prompt group at a time. This enables finer-grained zero-advantage skipping and reduces peak memory usage during scoring.
+
+```yaml
+trl:
+  streaming_partial_batch: true
+```
+
+##### Importance Sampling Correction
+
+When using async prefetch, completions are generated from a slightly older version of the model. Importance sampling (IS) correction adjusts the policy gradient to account for this distribution shift.
+
+```yaml
+trl:
+  vllm_importance_sampling_correction: true   # Enable IS correction
+  importance_sampling_level: token             # 'token' or 'sequence'
+  off_policy_mask_threshold: 0.5              # Mask sequences with IS ratio below this
+```
+
+- `importance_sampling_level: token` applies per-token IS ratios (recommended with Liger kernel)
+- `importance_sampling_level: sequence` applies per-sequence IS ratios
+- `off_policy_mask_threshold` masks out sequences where the IS ratio indicates they are too far off-policy
+
+##### Replay Buffer
+
+The replay buffer caches rollout groups that had learning signal (non-zero reward variance) and uses them to replace zero-signal groups in later batches.
+
+```yaml
+trl:
+  replay_buffer_size: 100       # Max cached groups (0 = disabled)
+  replay_recompute_logps: true  # Recompute log-probs for replayed data (recommended)
+```
+
+::: {.callout-note}
+When `replay_recompute_logps: true` (default), old log-probabilities are recomputed using the current model weights. This fixes the IS mismatch that would otherwise occur when replaying stale data.
+:::
+
+##### Deferred Re-rolling
+
+Failed prompts (where the model produces zero reward for all generations) are buffered and re-injected into later batches when the model may be better equipped to solve them.
+
+```yaml
+trl:
+  reroll_start_fraction: 0.5    # Start re-rolling after 50% of training
+  reroll_max_groups: 1          # Max groups to replace per batch
+```
+
+##### Zero-Advantage Batch Skipping
+
+When all advantages in a micro-batch are zero (no learning signal), the forward/backward pass is skipped entirely. This is enabled by default and logged as `skipped_zero_adv_batches=1`.
+
+```yaml
+trl:
+  skip_zero_advantage_batches: true   # default
+```
+
+##### Parallel Reward Workers
+
+Reward functions that use `signal.alarm()` (e.g., `math_verify`) must run in the main thread. Parallel reward workers use subprocesses to work around this limitation while enabling concurrent reward computation.
+
+```yaml
+trl:
+  reward_num_workers: 4         # Number of subprocess workers (1 = no parallelism)
+```
+
+##### Full Async GRPO Example
+
+```yaml
+base_model: Qwen/Qwen2.5-1.5B-Instruct
+
+vllm:
+    host: 0.0.0.0
+    port: 8000
+    gpu_memory_utilization: 0.35
+    dtype: auto
+
+adapter: lora
+lora_r: 32
+lora_alpha: 64
+lora_target_linear: true
+
+rl: grpo
+trl:
+  use_data_producer: true
+  use_vllm: true
+  async_prefetch: true
+  prefetch_depth: 1
+  vllm_sync_interval: 2
+  vllm_lora_sync: true
+  streaming_partial_batch: true
+  vllm_importance_sampling_correction: true
+  off_policy_mask_threshold: 0.5
+  importance_sampling_level: token
+  num_generations: 8
+  max_completion_length: 512
+  reward_funcs:
+    - rewards.accuracy_reward
+  reroll_start_fraction: 0.5
+  replay_buffer_size: 100
+  reward_num_workers: 4
+  skip_zero_advantage_batches: true
+
+datasets:
+  - path: AI-MO/NuminaMath-TIR
+    type: rewards.prompt_transform
+    split: train
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+max_steps: 500
+learning_rate: 1e-5
+bf16: true
+gradient_checkpointing: true
+```
+
+```bash
+# Terminal 1: Start vLLM on GPU 0
+CUDA_VISIBLE_DEVICES=0 axolotl vllm-serve config.yaml
+
+# Terminal 2: Train on GPU 1
+CUDA_VISIBLE_DEVICES=1 axolotl train config.yaml
+```
+
+##### Multi-GPU Async GRPO
+
+Async GRPO supports FSDP and DeepSpeed ZeRO-3 for multi-GPU training. vLLM runs on one GPU while training is distributed across the remaining GPUs.
+
+**FSDP:**
+
+```yaml
+fsdp:
+  - full_shard
+  - auto_wrap
+fsdp_config:
+  fsdp_transformer_layer_cls_to_wrap: Qwen2DecoderLayer
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+```
+
+**DeepSpeed ZeRO-3:**
+
+```yaml
+deepspeed: deepspeed_configs/zero3_bf16.json
+gradient_checkpointing_kwargs:
+  use_reentrant: true   # Required for ZeRO-3
+```
+
+```bash
+# Terminal 1: Start vLLM on GPU 0
+CUDA_VISIBLE_DEVICES=0 axolotl vllm-serve config.yaml
+
+# Terminal 2: Train on GPUs 0,1
+CUDA_VISIBLE_DEVICES=0,1 accelerate launch --num_processes 2 -m axolotl.cli.train config.yaml
+```
+
+::: {.callout-important}
+With multi-GPU async prefetch, only rank 0 generates completions in the background thread. Results are broadcast to all ranks on the main thread. This avoids FSDP/DeepSpeed collective deadlocks from unsynchronized background threads.
+:::
+
 ### GDPO
 
 GDPO (Group Reward-Decoupled Policy Optimization) extends GRPO for multi-reward training. It addresses the **reward advantage collapse** problem by normalizing each reward function independently before combining them.

examples/colab-notebooks/colab-axolotl-example.ipynb

@@ -40,7 +40,7 @@
     "%%capture\n",
     "# This step can take ~5-10 minutes to install dependencies\n",
     "!pip install --no-build-isolation axolotl[flash-attn]>=0.9.1\n",
-    "!pip install \"cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@e8ad129\""
+    "!pip install \"cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@63b15e6\""
    ]
   },
   {

examples/gemma3/gemma-3-1b-qlora.yml

@@ -1,8 +1,5 @@
 base_model: google/gemma-3-1b-it
 
-model_type: Gemma3ForCausalLM
-cls_model_config: Gemma3TextConfig
-
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
 
@@ -27,6 +24,11 @@ datasets:
 val_set_size: 0.0
 output_dir: ./outputs/out
 
+# Freeze vision tower
+unfrozen_parameters:
+  - ^model\.language_model\..*
+  - ^lm_head\..*
+
 adapter: qlora
 lora_r: 32
 lora_alpha: 16

examples/gemma3/gemma-3-270m-qlora.yml

@@ -1,8 +1,5 @@
 base_model: google/gemma-3-270m-it
 
-model_type: Gemma3ForCausalLM
-cls_model_config: Gemma3TextConfig
-
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
 
@@ -27,6 +24,11 @@ datasets:
 val_set_size: 0.0
 output_dir: ./outputs/out
 
+# Freeze vision tower
+unfrozen_parameters:
+  - ^model\.language_model\..*
+  - ^lm_head\..*
+
 adapter: qlora
 lora_r: 32
 lora_alpha: 16

examples/gemma3/gemma-3-4b-qlora.yml

@@ -1,9 +1,5 @@
 base_model: google/gemma-3-4b-it
 
-# Need to set else transformers tries to load vision too
-model_type: Gemma3ForCausalLM
-cls_model_config: Gemma3TextConfig
-
 load_in_4bit: true
 
 # gemma3 doesn't seem to play nice with ddp
@@ -24,6 +20,11 @@ dataset_prepared_path: last_run_prepared
 val_set_size: 0.01
 output_dir: ./outputs/out
 
+# Freeze vision tower
+unfrozen_parameters:
+  - ^model\.language_model\..*
+  - ^lm_head\..*
+
 adapter: qlora
 lora_model_dir:
 

examples/mistral4/README.md

@@ -0,0 +1,82 @@
+# Finetune Mistral Small 4 with Axolotl
+
+Mistral Small 4 is a 119B parameter (6.5B active) multimodal MoE model from MistralAI that unifies instruct, reasoning, and coding capabilities into a single model. It is available on HuggingFace at [Mistral-Small-4-119B-2603](https://huggingface.co/mistralai/Mistral-Small-4-119B-2603).
+
+Thanks to the team at MistralAI for giving us early access to prepare for this release.
+
+## Getting started
+
+1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html).
+
+2. Install [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy) to reduce training VRAM usage
+
+3. Install transformers from main
+
+  ```bash
+  pip install git+https://github.com/huggingface/transformers.git
+  ```
+
+4. Run one of the example configs:
+
+  ```bash
+  # text-only
+  axolotl train examples/mistral4/qlora-text.yml  # no experts ~69 GiB, experts ~93 GiB
+  axolotl train examples/mistral4/fft-text.yml
+
+  # text + vision
+  # run: wget https://huggingface.co/datasets/Nanobit/text-vision-2k-test/resolve/main/African_elephant.jpg
+  axolotl train examples/mistral4/qlora-vision.yml  # no experts ~68 GiB
+  axolotl train examples/mistral4/fft-vision.yml
+  ```
+
+Note: FFT configs provided as reference. Please adjust hyperparameters as needed.
+
+## Reasoning Effort
+
+The chat template supports a `reasoning_effort` variable to control the model's reasoning depth:
+
+- `"none"` — instruct mode (default)
+- `"high"` — reasoning mode with explicit thinking steps
+
+Pass it via `chat_template_kwargs` under your dataset config:
+
+```yaml
+datasets:
+  - path: your/dataset
+    type: chat_template
+    chat_template_kwargs:
+      reasoning_effort: high
+```
+
+## Thinking Support
+
+The chat template supports a `thinking` content type in assistant messages for training on reasoning traces (rendered as `[THINK]...[/THINK]` blocks).
+
+To use thinking datasets, add the `thinking` mapping via `message_property_mappings`:
+
+```yaml
+datasets:
+  - path: your/thinking-dataset
+    type: chat_template
+    message_property_mappings:
+      role: role
+      content: content
+      thinking: thinking
+    chat_template_kwargs:
+      reasoning_effort: high
+```
+
+See the [Magistral thinking guide](../magistral/think/README.md) for dataset format details.
+
+## Tips
+
+- Read more on how to load your own dataset at [docs](https://docs.axolotl.ai/docs/dataset_loading.html).
+- The text dataset format follows the OpenAI Messages format as seen [here](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#chat_template).
+- The vision model requires multi-modal dataset format as documented [here](https://docs.axolotl.ai/docs/multimodal.html#dataset-format).
+
+## Related Resources
+
+- [MistralAI Mistral Small 4 Blog](https://mistral.ai/news/mistral-small-4)
+- [Axolotl Docs](https://docs.axolotl.ai)
+- [Axolotl GitHub](https://github.com/axolotl-ai-cloud/axolotl)
+- [Axolotl Discord](https://discord.gg/7m9sfhzaf3)

examples/mistral4/fft-text.yml

@@ -0,0 +1,58 @@
+base_model: axolotl-ai-co/Mistral-Small-4-119B-2603-BF16
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+  - axolotl.integrations.kernels.KernelsPlugin
+use_kernels: true
+use_sonicmoe: true
+
+# only train language model layers, freeze vision tower
+unfrozen_parameters:
+  - model.language_model.*
+  - lm_head
+  - embed_tokens
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./outputs/out
+
+sequence_len: 2048
+sample_packing: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 2e-5
+
+bf16: true
+tf32: true
+
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+
+fsdp_version: 2
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: false
+  state_dict_type: FULL_STATE_DICT
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Mistral4DecoderLayer
+  reshard_after_forward: true
+  activation_checkpointing: true

examples/mistral4/fft-vision.yml

@@ -0,0 +1,57 @@
+base_model: axolotl-ai-co/Mistral-Small-4-119B-2603-BF16
+processor_type: AutoProcessor
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+  - axolotl.integrations.kernels.KernelsPlugin
+use_kernels: true
+use_sonicmoe: true
+
+# vision requirements
+skip_prepare_dataset: true
+remove_unused_columns: false
+sample_packing: false
+
+datasets:
+  - path: Nanobit/text-vision-2k-test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./outputs/out
+
+sequence_len: 2048
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 2e-5
+
+bf16: true
+tf32: true
+
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+
+fsdp_version: 2
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: false
+  state_dict_type: FULL_STATE_DICT
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Mistral4DecoderLayer
+  reshard_after_forward: true
+  activation_checkpointing: true

examples/mistral4/qlora-text.yml

@@ -0,0 +1,58 @@
+base_model: axolotl-ai-co/Mistral-Small-4-119B-2603-BF16
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+load_in_4bit: true
+quantize_moe_experts: true
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./outputs/out
+
+adapter: qlora
+
+sequence_len: 2048
+sample_packing: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules: 'model.language_model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
+
+# uncomment to train on expert layers
+# lora_target_parameters:
+#   - mlp.experts.gate_up_proj
+#   - mlp.experts.down_proj
+# lora_mlp_kernel: false
+# lora_qkv_kernel: false
+# lora_o_kernel: false
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: true
+tf32: true
+
+gradient_checkpointing: true
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0

examples/mistral4/qlora-vision.yml

@@ -0,0 +1,63 @@
+base_model: axolotl-ai-co/Mistral-Small-4-119B-2603-BF16
+processor_type: AutoProcessor
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+load_in_4bit: true
+quantize_moe_experts: true
+
+# vision chat template requirements
+skip_prepare_dataset: true
+remove_unused_columns: false
+sample_packing: false
+
+datasets:
+  - path: Nanobit/text-vision-2k-test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./outputs/out
+
+adapter: qlora
+
+sequence_len: 2048
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules: 'model.language_model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
+
+# uncomment to train on expert layers
+# lora_target_parameters:
+#   - mlp.experts.gate_up_proj
+#   - mlp.experts.down_proj
+# lora_mlp_kernel: false
+# lora_qkv_kernel: false
+# lora_o_kernel: false
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: true
+tf32: true
+
+gradient_checkpointing: true
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0

examples/nemotron/nemotron-mini-4b-qlora.yaml

@@ -0,0 +1,57 @@
+base_model: nvidia/Nemotron-Mini-4B-Instruct
+
+load_in_8bit: false
+load_in_4bit: true
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.1
+output_dir: ./outputs/nemotron-mini-4b-qlora
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 4096
+sample_packing: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_target_modules:
+  - q_proj
+  - k_proj
+  - v_proj
+  - o_proj
+  - up_proj
+  - down_proj
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+
+special_tokens:

examples/qwen3.5/122b-a10b-moe-qlora-fsdp.yaml

@@ -0,0 +1,84 @@
+base_model: Qwen/Qwen3.5-122B-A10B
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+strict: false
+
+chat_template: qwen3_5
+datasets:
+  - path: mlabonne/FineTome-100k
+    type: chat_template
+    split: train[:20%]
+    field_messages: conversations
+    message_property_mappings:
+      role: from
+      content: value
+val_set_size: 0.0
+output_dir: ./outputs/out
+dataset_prepared_path: last_run_prepared
+
+sequence_len: 2048
+sample_packing: true
+
+load_in_4bit: true
+quantize_moe_experts: true
+adapter: qlora
+lora_r: 16
+lora_alpha: 32
+lora_dropout: 0
+lora_target_modules:
+  - q_proj
+  - k_proj
+  - v_proj
+  - o_proj
+# Regex matching to target shared experts too
+# lora_target_modules: 'model\.(language_model\.)?layers\.[\d]+\.(mlp|self_attn)\.(shared_expert\.)?(up|down|gate|gate_up|q|k|v|o)_proj'
+
+# Target experts
+# lora_target_parameters:
+#   - mlp.experts.gate_up_proj
+#   - mlp.experts.down_proj
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_4bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: true
+
+lora_mlp_kernel: false
+lora_qkv_kernel: false
+lora_o_kernel: false
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:
+
+fsdp_config:
+  fsdp_version: 2
+  offload_params: true
+  cpu_ram_efficient_loading: false
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Qwen3_5MoeDecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  sharding_strategy: FULL_SHARD
+  reshard_after_forward: true
+  activation_checkpointing: true

examples/qwen3.5/122b-a10b-moe-qlora.yaml

@@ -32,7 +32,11 @@ lora_target_modules:
   - v_proj
   - o_proj
 
-#lora_target_parameters:
+# Regex matching to target shared experts too
+# lora_target_modules: 'model\.(language_model\.)?layers\.[\d]+\.(mlp|self_attn)\.(shared_expert\.)?(up|down|gate|gate_up|q|k|v|o)_proj'
+
+# Target experts
+# lora_target_parameters:
 #   - mlp.experts.gate_up_proj
 #   - mlp.experts.down_proj
 
@@ -52,7 +56,6 @@ learning_rate: 0.0002
 bf16: auto
 tf32: true
 
-
 lora_mlp_kernel: false
 lora_qkv_kernel: false
 lora_o_kernel: false

examples/qwen3.5/27b-fft.yaml

@@ -0,0 +1,59 @@
+base_model: Qwen/Qwen3.5-27B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+# Full fine-tune (FFT) of the text-only path of Qwen3.5-27B.
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+strict: false
+
+chat_template: qwen3_5
+datasets:
+  - path: mlabonne/FineTome-100k
+    type: chat_template
+    split: train[:20%]
+    field_messages: conversations
+    message_property_mappings:
+      role: from
+      content: value
+val_set_size: 0.0
+output_dir: ./outputs/out
+dataset_prepared_path: last_run_prepared
+
+sequence_len: 2048
+sample_packing: true
+
+# Freeze vision encoder
+unfrozen_parameters:
+  - model\.language_model\..*
+  - lm_head\..*
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: true
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

examples/qwen3.5/27b-qlora-fsdp.yaml

@@ -0,0 +1,81 @@
+base_model: Qwen/Qwen3.5-27B
+
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+strict: false
+
+chat_template: qwen3_5
+datasets:
+  - path: mlabonne/FineTome-100k
+    type: chat_template
+    split: train[:20%]
+    field_messages: conversations
+    message_property_mappings:
+      role: from
+      content: value
+val_set_size: 0.0
+output_dir: ./outputs/out
+dataset_prepared_path: last_run_prepared
+
+sequence_len: 2048
+sample_packing: true
+
+load_in_4bit: true
+adapter: qlora
+lora_r: 16
+lora_alpha: 32
+lora_target_modules:
+  - q_proj
+  - k_proj
+  - v_proj
+  - o_proj
+  - down_proj
+  - up_proj
+  # Uncomment below to also target the linear attention projections.
+  # These use separate in_proj_qkv / in_proj_z / out_proj (Qwen3.5-specific).
+  # - linear_attn.in_proj_qkv
+  # - linear_attn.in_proj_z
+  # - linear_attn.out_proj
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_4bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: true
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:
+
+fsdp_config:
+  fsdp_version: 2
+  offload_params: false
+  cpu_ram_efficient_loading: false
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Qwen3_5DecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  sharding_strategy: FULL_SHARD
+  reshard_after_forward: true
+  activation_checkpointing: true

examples/qwen3.5/27b-qlora.yaml

@@ -1,9 +1,7 @@
 base_model: Qwen/Qwen3.5-27B
+
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
-# Note: Qwen3.5 is an early-fusion VLM (image+text). This config fine-tunes
-# the text-only path. For multimodal (image+text) fine-tuning, add image
-# columns to your dataset following axolotl's multimodal dataset format.
 
 plugins:
   - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

examples/qwen3.5/35b-a3b-moe-qlora-fsdp.yaml

@@ -0,0 +1,85 @@
+base_model: Qwen/Qwen3.5-35B-A3B
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+strict: false
+
+chat_template: qwen3_5
+datasets:
+  - path: mlabonne/FineTome-100k
+    type: chat_template
+    split: train[:20%]
+    field_messages: conversations
+    message_property_mappings:
+      role: from
+      content: value
+val_set_size: 0.0
+output_dir: ./outputs/out
+dataset_prepared_path: last_run_prepared
+
+sequence_len: 2048
+sample_packing: true
+
+load_in_4bit: true
+quantize_moe_experts: true
+adapter: qlora
+lora_r: 16
+lora_alpha: 32
+lora_dropout: 0
+lora_target_modules:
+  - q_proj
+  - k_proj
+  - v_proj
+  - o_proj
+
+# Regex matching to target shared experts too
+# lora_target_modules: 'model\.(language_model\.)?layers\.[\d]+\.(mlp|self_attn)\.(shared_expert\.)?(up|down|gate|gate_up|q|k|v|o)_proj'
+
+# Target experts
+# lora_target_parameters:
+#   - mlp.experts.gate_up_proj
+#   - mlp.experts.down_proj
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_4bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: true
+
+lora_mlp_kernel: false
+lora_qkv_kernel: false
+lora_o_kernel: false
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:
+
+fsdp_config:
+  fsdp_version: 2
+  offload_params: true
+  cpu_ram_efficient_loading: false
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Qwen3_5MoeDecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  sharding_strategy: FULL_SHARD
+  reshard_after_forward: true
+  activation_checkpointing: true

examples/qwen3.5/35b-a3b-moe-qlora.yaml

@@ -32,7 +32,11 @@ lora_target_modules:
   - v_proj
   - o_proj
 
-#lora_target_parameters:
+# Regex matching to target shared experts too
+# lora_target_modules: 'model\.(language_model\.)?layers\.[\d]+\.(mlp|self_attn)\.(shared_expert\.)?(up|down|gate|gate_up|q|k|v|o)_proj'
+
+# Target experts
+# lora_target_parameters:
 #   - mlp.experts.gate_up_proj
 #   - mlp.experts.down_proj
 

examples/qwen3.5/9b-fft-vision.yaml

@@ -0,0 +1,49 @@
+base_model: Qwen/Qwen3.5-9B
+processor_type: AutoProcessor
+
+# Required for multimodal training
+skip_prepare_dataset: true
+remove_unused_columns: false
+sample_packing: false
+
+chat_template: qwen3_5
+datasets:
+  - path: HuggingFaceH4/llava-instruct-mix-vsft
+    type: chat_template
+    split: train[:1%]
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+sequence_len: 4096
+pad_to_sequence_len: false
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: true
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

examples/qwen3.5/9b-lora-vision.yaml

@@ -1,10 +1,6 @@
-base_model: Qwen/Qwen3.5-7B
+base_model: Qwen/Qwen3.5-9B
 processor_type: AutoProcessor
 
-# Qwen3.5-7B and above are early-fusion VLMs (Qwen3_5ForConditionalGeneration).
-# Vision and text tokens are processed together by the same transformer layers.
-# Note: Qwen3.5-2B is a text-only model — the smallest VLM is Qwen3.5-7B.
-
 # These 3 lines are required for vision/multimodal training
 skip_prepare_dataset: true
 remove_unused_columns: false
@@ -30,8 +26,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 # Targets the language model attention and MLP layers.
-# Qwen3.5 is early-fusion: all layers (including those seeing vision tokens) share
-# the same transformer stack, so standard attention targets work for both modalities.
 lora_target_modules:
   - q_proj
   - k_proj

examples/qwen3.5/README.md

@@ -1,15 +1,6 @@
 # Finetune Qwen3.5 with Axolotl
 
-[Qwen3.5](https://huggingface.co/collections/Qwen/qwen35-68452f3bc6e4b7cfb4e1c803) is a hybrid architecture model series combining Gated DeltaNet linear attention with standard Transformer attention. Models from 7B onwards are early-fusion vision-language models (`Qwen3_5ForConditionalGeneration`), meaning vision and text tokens are processed through the same transformer stack. The 2B variant is text-only.
-
-Available configs:
-
-| Config | Model | Type |
-|---|---|---|
-| `27b-qlora.yaml` | Qwen3.5-27B | Dense VLM, text-only path |
-| `35b-a3b-moe-qlora.yaml` | Qwen3.5-35B-A3B | MoE, text-only path |
-| `122b-a10b-moe-qlora.yaml` | Qwen3.5-122B-A10B | MoE, text-only path |
-| `7b-lora-vision.yaml` | Qwen3.5-7B | Vision+text (multimodal) |
+[Qwen3.5](https://huggingface.co/collections/Qwen/qwen35) is a hybrid architecture model series combining Gated DeltaNet linear attention with standard Transformer attention. All Qwen3.5 models are early-fusion vision-language models: dense variants use `Qwen3_5ForConditionalGeneration` and MoE variants use `Qwen3_5MoeForConditionalGeneration`.
 
 ## Getting started
 
@@ -18,35 +9,69 @@ Available configs:
 2. Install [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy) to reduce training VRAM usage.
 
 3. Install FLA for sample packing support with the Gated DeltaNet linear attention layers:
-```bash
-pip3 uninstall -y causal-conv1d && pip3 install flash-linear-attention==0.4.1
+  ```bash
+  pip3 uninstall -y causal-conv1d && pip3 install flash-linear-attention==0.4.1
+  ```
+  > FLA is required when `sample_packing: true`. Without it, training raises a `RuntimeError` on packed sequences. Vision configs use `sample_packing: false` so FLA is optional there.
+
+4. Pick any config from the table below and run:
+
+    ```bash
+    axolotl train examples/qwen3.5/<config>.yaml
+    ```
+
+Available configs:
+
+| Config | Model | Type | Peak VRAM |
+|---|---|---|---|
+| `9b-lora-vision.yaml` | Qwen3.5-9B | Vision+text LoRA, single GPU | — |
+| `9b-fft-vision.yaml` | Qwen3.5-9B | Vision+text FFT, single GPU | ~61 GiB |
+| `27b-qlora.yaml` | Qwen3.5-27B | Dense, text-only QLoRA | ~47 GiB |
+| `27b-fft.yaml` | Qwen3.5-27B | Dense, text-only FFT (vision frozen) | ~53 GiB |
+| `27b-qlora-fsdp.yaml` | Qwen3.5-27B | Dense, text-only QLoRA + FSDP2 | — |
+| `35b-a3b-moe-qlora.yaml` | Qwen3.5-35B-A3B | MoE, text-only QLoRA | — |
+| `35b-a3b-moe-qlora-fsdp.yaml` | Qwen3.5-35B-A3B | MoE, text-only QLoRA + FSDP2 | — |
+| `122b-a10b-moe-qlora.yaml` | Qwen3.5-122B-A10B | MoE, text-only QLoRA | — |
+| `122b-a10b-moe-qlora-fsdp.yaml` | Qwen3.5-122B-A10B | MoE, text-only QLoRA + FSDP2 | — |
+
+### Gated DeltaNet Linear Attention
+
+Qwen3.5 interleaves standard attention with Gated DeltaNet linear attention layers. To apply LoRA to them, add to `lora_target_modules`:
+
+```yaml
+lora_target_modules:
+  # ... standard projections ...
+  - linear_attn.in_proj_qkv
+  - linear_attn.in_proj_z
+  - linear_attn.out_proj
 ```
-> FLA is required when `sample_packing: true`. Without it, training raises a `RuntimeError` on packed sequences. Vision configs use `sample_packing: false` so FLA is optional there.
 
-4. Run a finetuning example:
+### Routed Experts (MoE)
 
-```bash
-# Dense 27B text-only (QLoRA, ~47 GiB VRAM with sample packing)
-axolotl train examples/qwen3.5/27b-qlora.yaml
+To apply LoRA to routed expert parameters, add `lora_target_parameters`:
 
-# MoE 35B-A3B text-only (QLoRA)
-axolotl train examples/qwen3.5/35b-a3b-moe-qlora.yaml
+```yaml
+lora_target_parameters:
+  - mlp.experts.gate_up_proj
+  - mlp.experts.down_proj
+#  - mlp.gate.weight  # router
+```
 
-# MoE 122B-A10B text-only (QLoRA)
-axolotl train examples/qwen3.5/122b-a10b-moe-qlora.yaml
+### Shared Experts (MoE)
 
-# 7B vision+text (LoRA, multimodal dataset)
-axolotl train examples/qwen3.5/7b-lora-vision.yaml
+Routed experts and shared experts both have `gate_up_proj`/`down_proj`, so a plain module name in `lora_target_modules` would match both. Use a regex to target only attention and shared expert projections, while `lora_target_parameters` above handles routed experts separately:
+
+```yaml
+lora_target_modules: 'model\.(language_model\.)?layers\.[\d]+\.(mlp|self_attn)\.(shared_expert\.)?(up|down|gate|gate_up|q|k|v|o)_proj'
 ```
 
 ### TIPS
 
-- For inference, you can experiment with `temperature: 0.7`, `top_p: 0.8`, `top_k: 20`, and `min_p: 0`.
-- You can run a full finetuning by removing `adapter: qlora` and `load_in_4bit: true`. See [Multi-GPU](#optimization-guides) below.
+- For inference hyp, please see the respective model card details.
+- You can run a full finetuning of smaller configs by removing `adapter: qlora` and `load_in_4bit: true`. See [Multi-GPU](#optimization-guides) below.
 - Read more on loading your own dataset at [docs](https://docs.axolotl.ai/docs/dataset_loading.html).
 - The dataset format follows the OpenAI Messages format as seen [here](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#chat_template).
-- For **multimodal** finetuning, set `processor_type: AutoProcessor`, `skip_prepare_dataset: true`, and `remove_unused_columns: false` as shown in `7b-lora-vision.yaml`.
-- The Gated DeltaNet linear attention layers (`linear_attn.*`) can optionally be added to `lora_target_modules` — they are commented out by default.
+- For **multimodal** finetuning, set `processor_type: AutoProcessor`, `skip_prepare_dataset: true`, and `remove_unused_columns: false` as shown in `9b-lora-vision.yaml`.
 
 ## Optimization Guides
 

pyproject.toml

@@ -61,5 +61,11 @@ skip-magic-trailing-comma = false
 line-ending = "auto"
 docstring-code-format = false
 
+[tool.pytest.ini_options]
+addopts = "-m 'not slow'"
+markers = [
+    "slow: marks tests as slow",
+]
+
 [tool.uv.extra-build-dependencies]
 axolotl = ["huggingface_hub"]

requirements.txt

@@ -75,4 +75,4 @@ axolotl-contribs-mit==0.0.6
 # telemetry
 posthog==6.7.11
 
-mistral-common==1.8.8
+mistral-common==1.10.0

scripts/cutcrossentropy_install.py

@@ -29,5 +29,5 @@ UV_PREFIX = "uv " if USE_UV else ""
 
 print(
     UNINSTALL_PREFIX
-    + f'{UV_PREFIX}pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@e8ad129"'
+    + f'{UV_PREFIX}pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@63b15e6"'
 )

setup.py

@@ -81,16 +81,23 @@ def parse_requirements(extras_require_map):
                     f"https://download.pytorch.org/whl/{torch_cuda_version}"
                 )
 
-            if (major, minor) >= (2, 9):
+            if (major, minor) >= (2, 10):
+                extras_require_map.pop("fbgemm-gpu")
+                extras_require_map["fbgemm-gpu"] = [
+                    "fbgemm-gpu==1.5.0",
+                    "fbgemm-gpu-genai==1.5.0",
+                ]
+                if not install_xformers:
+                    _install_requires.pop(_install_requires.index(xformers_version))
+                extras_require_map["vllm"] = ["vllm==0.17.1"]
+            elif (major, minor) >= (2, 9):
                 extras_require_map.pop("fbgemm-gpu")
                 extras_require_map["fbgemm-gpu"] = [
                     "fbgemm-gpu==1.4.0",
                     "fbgemm-gpu-genai==1.4.2",
                 ]
-                extras_require_map["vllm"] = ["vllm==0.11.1"]
                 if not install_xformers:
                     _install_requires.pop(_install_requires.index(xformers_version))
-                extras_require_map["vllm"] = ["vllm==0.13.0"]
                 if patch == 0:
                     extras_require_map["vllm"] = ["vllm==0.13.0"]
                 else:

src/axolotl/cli/checks.py

@@ -3,6 +3,7 @@
 import os
 from pathlib import Path
 
+import httpcore
 from accelerate.commands.config import config_args
 from huggingface_hub import HfApi
 from huggingface_hub.utils import LocalTokenNotFoundError
@@ -47,7 +48,7 @@ def check_user_token() -> bool:
             "Error verifying HuggingFace token. Remember to log in using `hf auth login` and get your access token from https://huggingface.co/settings/tokens if you want to use gated models or datasets."
         )
         return False
-    except HTTPError:
+    except (HTTPError, httpcore.ConnectError):
         LOG.warning(
             "Error accessing HuggingFace. This may be due to a network issue or rate limiting."
         )

src/axolotl/cli/cloud/modal_.py

@@ -90,9 +90,8 @@ class ModalCloud(Cloud):
         # grab the sha256 hash from docker hub for this image+tag
         # this ensures that we always get the latest image for this tag, even if it's already cached
         try:
-            manifest = subprocess.check_output(  # nosec B602
-                f"docker manifest inspect {docker_image}",
-                shell=True,
+            manifest = subprocess.check_output(  # nosec
+                ["docker", "manifest", "inspect", docker_image],
             ).decode("utf-8")
             sha256_hash = json.loads(manifest)["manifests"][0]["digest"]
         except subprocess.CalledProcessError:

src/axolotl/cli/config.py

@@ -11,7 +11,7 @@ from urllib.parse import urlparse
 import requests
 import torch
 import yaml
-from transformers.utils import is_torch_bf16_gpu_available
+from transformers.utils import is_torch_bf16_gpu_available, is_torch_tf32_available
 
 from axolotl.integrations.base import PluginManager
 from axolotl.telemetry.errors import send_errors
@@ -300,7 +300,7 @@ def load_cfg(
     try:
         device_props = torch.cuda.get_device_properties("cuda")
         gpu_version = "sm_" + str(device_props.major) + str(device_props.minor)
-    except:
+    except (RuntimeError, AssertionError):
         gpu_version = None
 
     prepare_plugins(cfg)
@@ -310,6 +310,7 @@ def load_cfg(
         capabilities={
             "bf16": is_torch_bf16_gpu_available(),
             "fp8": compute_supports_fp8(),
+            "tf32": is_torch_tf32_available(),
             "n_gpu": int(os.environ.get("WORLD_SIZE", 1)),
             "compute_capability": gpu_version,
         },

src/axolotl/cli/merge_sharded_fsdp_weights.py

@@ -196,12 +196,10 @@ def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs):
     state.wait_for_everyone()
     LOG.info(
         f"FSDP SHARDED_STATE_DICT weights successfully merged to: {output_path}",
-        main_process_only=True,
     )
     LOG.info(
         "Merged weights are only the safetensors and doesn't include the model configuration "
         f"or tokenizer which may be found in {parsed_cfg.output_dir}.",
-        main_process_only=True,
     )
 
 

src/axolotl/cli/vllm_serve.py

@@ -38,7 +38,18 @@ def do_vllm_serve(
     cfg = load_cfg(config)
     model = cfg.base_model
 
-    serve_module = cli_args.get("serve_module", "trl.scripts.vllm_serve")
+    # Determine serve module: explicit CLI/config > auto-select from vllm_lora_sync > default
+    serve_module = cli_args.get("serve_module") or getattr(
+        cfg.vllm, "serve_module", None
+    )
+    if (
+        serve_module is None
+        and getattr(cfg, "trl", None)
+        and getattr(cfg.trl, "vllm_lora_sync", False)
+    ):
+        serve_module = "axolotl.scripts.vllm_serve_lora"
+    if serve_module is None:
+        serve_module = "trl.scripts.vllm_serve"
     vllm_serve_main = __import__(serve_module, fromlist=["main"]).main
     tensor_parallel_size = 1
     data_parallel_size = 1
@@ -68,7 +79,7 @@ def do_vllm_serve(
         cli_args.get("enable_reasoning") or cfg.vllm.enable_reasoning or False
     )
 
-    vllm_script_args = AxolotlScriptArguments(
+    base_kwargs = dict(
         model=model,
         tensor_parallel_size=tensor_parallel_size,
         data_parallel_size=data_parallel_size,
@@ -78,7 +89,21 @@ def do_vllm_serve(
         dtype=dtype,
         max_model_len=max_model_len,
         enable_prefix_caching=enable_prefix_caching,
-        reasoning_parser=reasoning_parser,
-        enable_reasoning=enable_reasoning,
     )
+
+    # Use LoRAScriptArguments when serving with native LoRA support
+    if serve_module == "axolotl.scripts.vllm_serve_lora":
+        from axolotl.scripts.vllm_serve_lora import LoRAScriptArguments
+
+        lora_kwargs = {}
+        if hasattr(cfg, "lora_r") and cfg.lora_r:
+            lora_kwargs["max_lora_rank"] = cfg.lora_r
+        vllm_script_args = LoRAScriptArguments(**base_kwargs, **lora_kwargs)
+    else:
+        vllm_script_args = AxolotlScriptArguments(
+            **base_kwargs,
+            reasoning_parser=reasoning_parser,
+            enable_reasoning=enable_reasoning,
+        )
+
     vllm_serve_main(vllm_script_args)

src/axolotl/common/architectures.py

@@ -16,6 +16,7 @@ MOE_ARCH_BLOCK = {
     "qwen3_vl_moe": "Qwen3VLMoeTextSparseMoeBlock",
     "deepseek_v2": "DeepseekV2MoE",
     "deepseek_v3": "DeepseekV3MoE",
+    "mistral4": "Mistral4MoE",
     "gpt_oss": "GptOssDecoderLayer",
     "lfm2_moe": "Lfm2MoeSparseMoeBlock",
     "afmoe": "AfmoeMoE",

src/axolotl/convert.py

@@ -67,7 +67,7 @@ class JsonToJsonlConverter:
         self.json_parser = json_parser
         self.jsonl_serializer = jsonl_serializer
 
-    def convert(self, input_file_path, output_file_path):
+    def convert(self, input_file_path):
         content = self.file_reader.read(input_file_path)
         data = self.json_parser.parse(content)
         # data = [r for r in data if r["conversations"]]  # vicuna cleaned has rows with empty conversations

src/axolotl/core/builders/base.py

@@ -250,7 +250,7 @@ class TrainerBuilderBase(abc.ABC):
 
     def _configure_precision_settings(self, training_args_kwargs: dict):
         training_args_kwargs["fp16"] = (self.cfg.fp16 and not self.cfg.bf16) or False
-        training_args_kwargs["tf32"] = self.cfg.tf32
+        training_args_kwargs["tf32"] = True if self.cfg.tf32 is True else False
         if self.cfg.bf16 == "full":
             training_args_kwargs["bf16_full_eval"] = True
         else:
@@ -353,6 +353,30 @@ class TrainerBuilderBase(abc.ABC):
                 adam_kwargs["eps"] = (eps1, eps2)
 
                 optimizer_kwargs.update(adam_kwargs)
+            elif self.cfg.optimizer == "flash_adamw":
+                from flashoptim import FlashAdamW
+
+                optimizer_cls = FlashAdamW
+                optimizer_kwargs.update(adam_kwargs)
+            elif self.cfg.optimizer == "flash_adam":
+                from flashoptim import FlashAdam
+
+                optimizer_cls = FlashAdam
+                optimizer_kwargs.update(adam_kwargs)
+            elif self.cfg.optimizer == "flash_sgd":
+                from flashoptim import FlashSGD
+
+                optimizer_cls = FlashSGD
+            elif self.cfg.optimizer == "flash_sgdw":
+                from flashoptim import FlashSGDW
+
+                optimizer_cls = FlashSGDW
+            elif self.cfg.optimizer == "flash_lion":
+                from flashoptim import FlashLion
+
+                optimizer_cls = FlashLion
+                if "betas" in adam_kwargs:
+                    optimizer_kwargs["betas"] = adam_kwargs["betas"]
             else:
                 raise ValueError(
                     f"Unhandled optimizer: {self.cfg.optimizer}. Please raise an Issue."
@@ -484,6 +508,8 @@ class TrainerBuilderBase(abc.ABC):
             training_args_kwargs["accelerator_config"] = AcceleratorConfig()
 
     def _configure_gradient_checkpointing(self, training_args_kwargs: dict):
+        if self.cfg.layer_offloading:
+            training_args_kwargs["layer_offloading"] = True
         if self.cfg.activation_offloading is True:
             # don't use the HF gradient checkpointing, manually wrap
             training_args_kwargs["gradient_checkpointing"] = False

src/axolotl/core/builders/causal.py

@@ -421,6 +421,13 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             trainer_kwargs["dataset_tags"] = [
                 d["path"] for d in self.cfg.datasets if not Path(d["path"]).is_dir()
             ]
+        # TRL's RewardTrainer validates num_labels=1 on pre-loaded models; ensure the
+        # config reflects this regardless of how the model was instantiated.
+        if (
+            self.cfg.reward_model
+            and getattr(self.model.config, "num_labels", None) != 1
+        ):
+            self.model.config.num_labels = 1
         trainer = trainer_cls(
             model=self.model,
             train_dataset=self.train_dataset,

src/axolotl/core/builders/rl.py

@@ -54,8 +54,16 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
         if self.cfg.rl in {RLType.GRPO, RLType.GDPO}:
             from axolotl.core.trainers.grpo import GRPOStrategy
 
+            async_grpo = bool(
+                self.cfg.trl
+                and (
+                    getattr(self.cfg.trl, "async_prefetch", False)
+                    or getattr(self.cfg.trl, "use_data_producer", False)
+                )
+            )
             trainer_cls = GRPOStrategy.get_trainer_class(
-                sequence_parallel=self.cfg.context_parallel_size > 1
+                sequence_parallel=self.cfg.context_parallel_size > 1,
+                async_grpo=async_grpo,
             )
             trainer_cls_args.extend(GRPOStrategy.set_trainer_args(self.cfg))
             trainer_kwargs.update(GRPOStrategy.set_trainer_kwargs(self.cfg))
@@ -151,7 +159,16 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
         elif self.cfg.rl in {RLType.GRPO, RLType.GDPO}:
             from axolotl.core.trainers.grpo import GRPOStrategy
 
-            training_args_cls = GRPOStrategy.get_training_args_class()
+            async_grpo = bool(
+                self.cfg.trl
+                and (
+                    getattr(self.cfg.trl, "async_prefetch", False)
+                    or getattr(self.cfg.trl, "use_data_producer", False)
+                )
+            )
+            training_args_cls = GRPOStrategy.get_training_args_class(
+                async_grpo=async_grpo
+            )
             training_args_kwargs.update(GRPOStrategy.set_training_args_kwargs(self.cfg))
             blocklist_args_kwargs = GRPOStrategy.get_blocklist_args_kwargs()
             if self.cfg.rl is RLType.GDPO:
@@ -191,7 +208,11 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
 
         if self.eval_dataset:
             trainer_kwargs["eval_dataset"] = self.eval_dataset
-        if self.cfg.adapter and self.peft_config and self.cfg.rl is not RLType.GRPO:
+        if (
+            self.cfg.adapter
+            and self.peft_config
+            and self.cfg.rl not in (RLType.GRPO, RLType.ORPO)
+        ):
             trainer_kwargs["peft_config"] = self.peft_config
         if self.cfg.precompute_ref_log_probs is not None:
             trainer_kwargs["precompute_ref_log_probs"] = (
@@ -217,13 +238,36 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             trainer_kwargs, trainer_cls
         )
 
-        trainer = trainer_cls(
-            *trainer_cls_args,
-            args=training_args,
-            train_dataset=self.train_dataset,
-            callbacks=self.get_callbacks(),
-            **trainer_kwargs,
-        )
+        # Allow FP8-quantized models to be fine-tuned with LoRA adapters.
+        # transformers' validate_quantization_for_training blocks FP8 because
+        # hf_quantizer.is_trainable is False, but LoRA only trains the adapters
+        # (base weights stay frozen in FP8).
+        _orig_validate_quant = None
+        if (
+            self.cfg.adapter
+            and hasattr(self.model, "is_quantized")
+            and self.model.is_quantized
+        ):
+            import transformers.trainer as _trainer_module
+
+            _orig_validate_quant = _trainer_module.validate_quantization_for_training
+            _trainer_module.validate_quantization_for_training = lambda model: None
+
+        try:
+            trainer = trainer_cls(
+                *trainer_cls_args,
+                args=training_args,
+                train_dataset=self.train_dataset,
+                callbacks=self.get_callbacks(),
+                **trainer_kwargs,
+            )
+        finally:
+            if _orig_validate_quant is not None:
+                import transformers.trainer as _trainer_module
+
+                _trainer_module.validate_quantization_for_training = (
+                    _orig_validate_quant
+                )
         if self.cfg.fsdp_config or self.cfg.fsdp:
             ensure_dtype(trainer.model, dtype=self.cfg.torch_dtype)
             if self.cfg.rl in [RLType.DPO, RLType.IPO] and trainer.ref_model:

src/axolotl/core/trainers/base.py

@@ -29,10 +29,12 @@ from transformers.utils import SAFE_WEIGHTS_NAME, is_peft_available
 from trl.experimental.utils import pad_to_length
 from typing_extensions import override
 
+from axolotl.core.trainers.constants import TOKENS_STATE_FILE
 from axolotl.core.trainers.mixins import (
     ActivationOffloadingMixin,
     CheckpointSaveMixin,
     DistributedParallelMixin,
+    LayerOffloadingMixin,
     OptimizerMixin,
     PackingMixin,
     RngLoaderMixin,
@@ -51,8 +53,6 @@ from axolotl.utils.samplers import MultipackBatchSampler, get_dataset_lengths
 
 LOG = get_logger(__name__)
 
-TOKENS_STATE_FILE = "tokens_state."
-
 REDUCTION_FNS = {
     "mean": torch.mean,
     "min": torch.min,
@@ -67,6 +67,7 @@ class AxolotlTrainer(
     OptimizerMixin,
     RngLoaderMixin,
     CheckpointSaveMixin,
+    LayerOffloadingMixin,
     ActivationOffloadingMixin,
     DistributedParallelMixin,
     Trainer,

src/axolotl/core/trainers/constants.py

@@ -0,0 +1 @@
+TOKENS_STATE_FILE = "tokens_state.json"

src/axolotl/core/trainers/dpo/args.py

@@ -2,7 +2,8 @@
 Axolotl specific DPO args
 """
 
-from dataclasses import dataclass
+from dataclasses import dataclass, field
+from typing import Optional
 
 from trl import DPOConfig
 
@@ -16,3 +17,4 @@ class AxolotlDPOConfig(AxolotlTrainingMixins, DPOConfig):
     """
 
     dpo_norm_loss: bool | None = False
+    rpo_alpha: Optional[float] = field(default=None)

src/axolotl/core/trainers/grpo/__init__.py

@@ -9,8 +9,9 @@ from huggingface_hub import snapshot_download
 from requests import HTTPError
 from trl.trainer.grpo_trainer import RewardFunc
 
-from axolotl.core.trainers.grpo.args import AxolotlGRPOConfig
+from axolotl.core.trainers.grpo.args import AxolotlAsyncGRPOConfig, AxolotlGRPOConfig
 from axolotl.core.trainers.grpo.trainer import (
+    AxolotlAsyncGRPOTrainer,
     AxolotlGRPOSequenceParallelTrainer,
     AxolotlGRPOTrainer,
 )
@@ -27,14 +28,31 @@ class GRPOStrategy:
 
     @classmethod
     def get_trainer_class(
-        cls, sequence_parallel: bool
-    ) -> type[AxolotlGRPOTrainer] | type[AxolotlGRPOSequenceParallelTrainer]:
+        cls,
+        sequence_parallel: bool,
+        async_grpo: bool = False,
+    ) -> (
+        type[AxolotlGRPOTrainer]
+        | type[AxolotlGRPOSequenceParallelTrainer]
+        | type[AxolotlAsyncGRPOTrainer]
+    ):
+        if sequence_parallel and async_grpo:
+            raise ValueError(
+                "sequence_parallel and async_grpo cannot both be enabled. "
+                "Disable one of context_parallel_size > 1 or async_prefetch/use_data_producer."
+            )
         if sequence_parallel:
             return AxolotlGRPOSequenceParallelTrainer
+        if async_grpo:
+            return AxolotlAsyncGRPOTrainer
         return AxolotlGRPOTrainer
 
     @classmethod
-    def get_training_args_class(cls) -> type[AxolotlGRPOConfig]:
+    def get_training_args_class(
+        cls, async_grpo: bool = False
+    ) -> type[AxolotlGRPOConfig] | type[AxolotlAsyncGRPOConfig]:
+        if async_grpo:
+            return AxolotlAsyncGRPOConfig
         return AxolotlGRPOConfig
 
     @classmethod
@@ -124,13 +142,63 @@ class GRPOStrategy:
             grpo_args_kwargs["epsilon_high"] = trl.epsilon_high
 
         if trl.use_liger_loss is not None:
-            grpo_args_kwargs["use_liger_loss"] = trl.use_liger_loss
+            grpo_args_kwargs["use_liger_kernel"] = trl.use_liger_loss
 
         if trl.multi_objective_aggregation is not None:
             grpo_args_kwargs["multi_objective_aggregation"] = (
                 trl.multi_objective_aggregation
             )
 
+        # Async GRPO fields
+        if getattr(trl, "use_data_producer", None) is not None:
+            grpo_args_kwargs["use_data_producer"] = trl.use_data_producer
+        if getattr(trl, "async_prefetch", None) is not None:
+            grpo_args_kwargs["async_prefetch"] = trl.async_prefetch
+        if getattr(trl, "prefetch_depth", None) is not None:
+            grpo_args_kwargs["prefetch_depth"] = trl.prefetch_depth
+        if getattr(trl, "vllm_sync_interval", None) is not None:
+            grpo_args_kwargs["vllm_sync_interval"] = trl.vllm_sync_interval
+        if getattr(trl, "streaming_partial_batch", None) is not None:
+            grpo_args_kwargs["streaming_partial_batch"] = trl.streaming_partial_batch
+        if getattr(trl, "streaming_min_groups", None) is not None:
+            grpo_args_kwargs["streaming_min_groups"] = trl.streaming_min_groups
+        if getattr(trl, "vllm_importance_sampling_correction", None) is not None:
+            grpo_args_kwargs["vllm_importance_sampling_correction"] = (
+                trl.vllm_importance_sampling_correction
+            )
+        if getattr(trl, "vllm_importance_sampling_mode", None) is not None:
+            grpo_args_kwargs["vllm_importance_sampling_mode"] = (
+                trl.vllm_importance_sampling_mode
+            )
+        if getattr(trl, "vllm_importance_sampling_cap", None) is not None:
+            grpo_args_kwargs["vllm_importance_sampling_cap"] = (
+                trl.vllm_importance_sampling_cap
+            )
+        if getattr(trl, "off_policy_mask_threshold", None) is not None:
+            grpo_args_kwargs["off_policy_mask_threshold"] = (
+                trl.off_policy_mask_threshold
+            )
+        if getattr(trl, "use_bias_correction_kl", None) is not None:
+            grpo_args_kwargs["use_bias_correction_kl"] = trl.use_bias_correction_kl
+
+        # Fast Async GRPO fields
+        if getattr(trl, "reward_num_workers", None) is not None:
+            grpo_args_kwargs["reward_num_workers"] = trl.reward_num_workers
+        if getattr(trl, "replay_buffer_size", None) is not None:
+            grpo_args_kwargs["replay_buffer_size"] = trl.replay_buffer_size
+        if getattr(trl, "replay_recompute_logps", None) is not None:
+            grpo_args_kwargs["replay_recompute_logps"] = trl.replay_recompute_logps
+        if getattr(trl, "reroll_start_fraction", None) is not None:
+            grpo_args_kwargs["reroll_start_fraction"] = trl.reroll_start_fraction
+        if getattr(trl, "reroll_max_groups", None) is not None:
+            grpo_args_kwargs["reroll_max_groups"] = trl.reroll_max_groups
+        if getattr(trl, "skip_zero_advantage_batches", None) is not None:
+            grpo_args_kwargs["skip_zero_advantage_batches"] = (
+                trl.skip_zero_advantage_batches
+            )
+        if getattr(trl, "vllm_lora_sync", None) is not None:
+            grpo_args_kwargs["vllm_lora_sync"] = trl.vllm_lora_sync
+
         return grpo_args_kwargs
 
     @classmethod

src/axolotl/core/trainers/grpo/args.py

@@ -6,6 +6,7 @@ from dataclasses import dataclass
 
 from trl import GRPOConfig
 
+from axolotl.core.trainers.grpo.fast_async_trainer import FastAsyncGRPOConfig
 from axolotl.core.training_args import AxolotlTrainingMixins
 
 
@@ -14,3 +15,10 @@ class AxolotlGRPOConfig(AxolotlTrainingMixins, GRPOConfig):
     """Axolotl GRPO Config for GRPO training"""
 
     context_parallel_size: int | None = None
+
+
+@dataclass
+class AxolotlAsyncGRPOConfig(AxolotlTrainingMixins, FastAsyncGRPOConfig):
+    """Axolotl Async GRPO Config — adds async prefetch, streaming scoring, and IS correction."""
+
+    context_parallel_size: int | None = None

src/axolotl/core/trainers/grpo/fast_async_trainer.py

@@ -0,0 +1,768 @@
+# Copyright 2020-2026 Axolotl AI. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Experimental GRPO extensions: parallel reward workers, replay buffer,
+deferred re-roll, and zero-advantage skipping.
+
+These features are built as subclasses of GRPOTrainer and GRPODataProducer,
+using the hook system (_compute_rewards_for_batch, _post_advantage_hook,
+_pre_produce_hook) defined in the base classes.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+import threading
+from dataclasses import dataclass, field
+
+import torch
+from torch import nn
+from trl import GRPOTrainer
+
+from axolotl.core.trainers.grpo.async_trainer import (
+    AsyncGRPOConfig,
+    AsyncGRPOTrainer,
+    GRPODataProducer,
+)
+from axolotl.core.trainers.grpo.replay_buffer import ReplayBuffer
+
+logger = logging.getLogger(__name__)
+
+
+# ---------------------------------------------------------------------------
+# Extended config
+# ---------------------------------------------------------------------------
+
+
+@dataclass
+class FastAsyncGRPOConfig(AsyncGRPOConfig):
+    """GRPOConfig with additional experimental parameters."""
+
+    reward_num_workers: int = field(
+        default=1,
+        metadata={
+            "help": "Number of persistent subprocess workers for parallel reward computation. Each worker has its "
+            "own main thread so signal.alarm() (used by math_verify) works correctly. Work is sharded across "
+            "workers by prompt groups. Only used with use_data_producer=True and non-nn.Module reward functions."
+        },
+    )
+    replay_buffer_size: int = field(
+        default=0,
+        metadata={
+            "help": "[Experimental, disabled by default] Size of the replay buffer for storing high-signal rollout "
+            "groups. When > 0, groups with reward variance are cached and used to replace zero-signal groups "
+            "(where all rewards are identical). Set to 0 to disable. Only used with use_data_producer=True."
+        },
+    )
+    replay_recompute_logps: bool = field(
+        default=True,
+        metadata={
+            "help": "When True (default), recompute old_per_token_logps for replayed groups using the current "
+            "training model. This fixes the importance sampling mismatch that occurs when replaying stale data. "
+            "Only relevant when replay_buffer_size > 0."
+        },
+    )
+    reroll_start_fraction: float = field(
+        default=0.5,
+        metadata={
+            "help": "Fraction of total training steps after which deferred re-rolling begins. Zero-signal prompts "
+            "(where all rewards in a group are identical) are buffered and re-injected into later batches when the "
+            "model is more likely to solve them. Set to 1.0 to disable. Only used with use_data_producer=True."
+        },
+    )
+    reroll_max_groups: int = field(
+        default=1,
+        metadata={
+            "help": "Maximum number of prompt groups to replace with re-roll candidates per batch. Higher values "
+            "increase data utilization but reduce prompt diversity. Only used with use_data_producer=True."
+        },
+    )
+    skip_zero_advantage_batches: bool = field(
+        default=True,
+        metadata={
+            "help": "When True, skip gradient computation for micro-batches where all advantages are zero (no learning "
+            "signal). This avoids the forward/backward pass entirely when no learning signal is present. The step is "
+            "logged with skipped_zero_adv_batches=1 for monitoring."
+        },
+    )
+    vllm_lora_sync: bool = field(
+        default=False,
+        metadata={
+            "help": "When True, sync LoRA adapter weights to vLLM via filesystem instead of merging into base model "
+            "and NCCL-broadcasting all parameters. vLLM loads the adapter natively using Punica kernels. "
+            "Requires vllm_serve_lora serve module (auto-selected when this is True). "
+            "Syncs only LoRA adapter weights (much smaller) vs full merged model. Legacy merge behavior is used when False."
+        },
+    )
+
+
+# ---------------------------------------------------------------------------
+# Extended data producer with re-roll injection
+# ---------------------------------------------------------------------------
+
+
+class RerollDataProducer(GRPODataProducer):
+    """GRPODataProducer that injects re-roll candidates into prompt batches.
+
+    Reads from the trainer's ``_reroll_buffer`` (populated by
+    ``GRPOExperimentalTrainer._post_advantage_hook``) and replaces the
+    last N prompt groups with previously-failed prompts.
+    """
+
+    def _pre_produce_hook(self, inputs: list, global_step: int) -> list:
+        trainer = self._trainer
+        reroll_buf = getattr(trainer, "_reroll_buffer", None)
+        reroll_lock = getattr(trainer, "_reroll_lock", None)
+        if reroll_buf is None or reroll_lock is None:
+            return inputs
+
+        max_steps = getattr(trainer.args, "max_steps", -1)
+        start_frac = getattr(trainer.args, "reroll_start_fraction", 1.0)
+        max_groups = getattr(trainer.args, "reroll_max_groups", 1)
+        reroll_start_step = (
+            max(1, int(max_steps * start_frac)) if max_steps > 0 else float("inf")
+        )
+
+        if global_step < reroll_start_step:
+            return inputs
+
+        with reroll_lock:
+            n_to_take = min(max_groups, len(reroll_buf))
+            reroll_prompts = [reroll_buf.pop(0) for _ in range(n_to_take)]
+
+        if reroll_prompts:
+            num_gen = self._num_generations
+            n_groups = len(inputs) // num_gen
+            for i, reroll_prompt in enumerate(reroll_prompts):
+                group_idx = n_groups - 1 - i
+                if group_idx < 0:
+                    break
+                start = group_idx * num_gen
+                for j in range(num_gen):
+                    inputs[start + j] = reroll_prompt
+            logger.info(
+                f"[REROLL] Step {global_step}: replaced {len(reroll_prompts)}/{n_groups} prompt groups "
+                f"with deferred re-roll candidates ({len(reroll_buf)} remaining)"
+            )
+
+        return inputs
+
+
+# ---------------------------------------------------------------------------
+# Persistent reward subprocess pool
+# ---------------------------------------------------------------------------
+
+
+def _persistent_reward_worker(conn):
+    """Long-lived reward worker. Receives work items, returns results."""
+    while True:
+        try:
+            msg = conn.recv()
+        except EOFError:
+            break
+        if msg is None:  # Shutdown signal
+            break
+        (
+            reward_funcs,
+            prompts,
+            completions,
+            completion_ids_list,
+            inputs,
+            reward_func_names,
+        ) = msg
+        try:
+            keys = [
+                key
+                for key in inputs[0]
+                if key not in ["prompt", "completion", "completion_ids"]
+            ]
+            reward_kwargs = {key: [example[key] for example in inputs] for key in keys}
+            results = []
+            for reward_func, _reward_func_name in zip(
+                reward_funcs, reward_func_names, strict=True
+            ):
+                output = reward_func(
+                    prompts=prompts,
+                    completions=completions,
+                    completion_ids=completion_ids_list,
+                    **reward_kwargs,
+                )
+                results.append(
+                    [float(r) if r is not None else float("nan") for r in output]
+                )
+            conn.send(results)
+        except Exception:
+            conn.send(None)
+
+
+# ---------------------------------------------------------------------------
+# Extended trainer
+# ---------------------------------------------------------------------------
+
+
+class FastAsyncGRPOTrainer(AsyncGRPOTrainer):
+    """GRPOTrainer with experimental extensions.
+
+    Adds:
+    - Parallel reward subprocess workers (``reward_num_workers``)
+    - Replay buffer for high-signal group reuse (``replay_buffer_size``)
+    - Deferred re-roll of failed prompts (``reroll_start_fraction``)
+    - Zero-advantage micro-batch skipping
+    """
+
+    def __init__(self, *args, **kwargs):
+        # These must be initialized before super().__init__() because
+        # _create_data_producer (called during super().__init__) needs them.
+        self._reroll_buffer: list = []
+        self._reroll_lock = threading.Lock()
+
+        # Temporarily suppress the base class's Liger + OPSM validation check,
+        # since this subclass supports it via a custom compute_liger_loss override.
+        grpo_args = kwargs.get("args")
+        if grpo_args is None:
+            for a in args:
+                if hasattr(a, "off_policy_mask_threshold"):
+                    grpo_args = a
+                    break
+        saved_threshold = None
+        if grpo_args is not None and getattr(grpo_args, "use_liger_kernel", False):
+            saved_threshold = grpo_args.off_policy_mask_threshold
+            grpo_args.off_policy_mask_threshold = None
+
+        super().__init__(*args, **kwargs)
+
+        if saved_threshold is not None:
+            grpo_args.off_policy_mask_threshold = saved_threshold
+            self.off_policy_mask_threshold = saved_threshold
+
+        # Replay buffer
+        if getattr(self.args, "replay_buffer_size", 0) > 0:
+            self._replay_buffer = ReplayBuffer(max_size=self.args.replay_buffer_size)
+        else:
+            self._replay_buffer = None
+        self._replay_recompute_logps = getattr(
+            self.args, "replay_recompute_logps", True
+        )
+
+        # Reward worker pool (lazy-initialized)
+        self._reward_workers = None
+
+    # -- Factory override: use RerollDataProducer ----------------------------
+
+    def _create_data_producer(self, args, train_dataset):
+        """Override to use RerollDataProducer for re-roll prompt injection."""
+        from axolotl.core.trainers.grpo.async_trainer import (
+            AsyncDataProducer,
+            ProducerConfig,
+        )
+
+        producer_config = ProducerConfig(
+            mini_epochs=args.num_iterations,
+            max_rollouts=None,
+            eval_during_produce=False,
+            empty_cache_before_produce=True,
+            empty_cache_after_produce=True,
+            async_prefetch=args.async_prefetch,
+            prefetch_depth=args.prefetch_depth,
+        )
+        data_producer = RerollDataProducer(
+            config=producer_config,
+            prompt_dataset=train_dataset,
+            num_generations=self.num_generations,
+            generation_batch_size=args.generation_batch_size,
+            train_batch_size=args.per_device_train_batch_size,
+            steps_per_generation=args.steps_per_generation,
+            shuffle_dataset=self.shuffle_dataset,
+            seed=args.seed,
+        )
+        data_producer.set_trainer(self)
+        if args.async_prefetch:
+            data_producer = AsyncDataProducer(
+                data_producer,
+                background_produce_kwargs={"skip_policy_logps": True},
+            )
+        return data_producer
+
+    # -- Reward worker pool --------------------------------------------------
+
+    def _get_reward_workers(self):
+        """Return a list of persistent reward worker subprocesses (lazy-initialized)."""
+        import multiprocessing as _mp
+
+        num_workers = getattr(self.args, "reward_num_workers", 1)
+        if num_workers < 1:
+            num_workers = 1
+
+        if self._reward_workers is not None:
+            alive = all(proc.is_alive() for conn, proc in self._reward_workers)
+            if alive and len(self._reward_workers) == num_workers:
+                return self._reward_workers
+            self._shutdown_reward_workers()
+
+        workers = []
+        for _ in range(num_workers):
+            parent_conn, child_conn = _mp.Pipe()
+            proc = _mp.Process(
+                target=_persistent_reward_worker, args=(child_conn,), daemon=True
+            )
+            proc.start()
+            child_conn.close()
+            workers.append((parent_conn, proc))
+
+        self._reward_workers = workers
+        return workers
+
+    def _shutdown_reward_workers(self):
+        """Shut down all persistent reward workers."""
+        if self._reward_workers is None:
+            return
+        for conn, proc in self._reward_workers:
+            try:
+                conn.send(None)
+                proc.join(timeout=5)
+            except Exception:
+                pass
+            try:
+                conn.close()
+            except Exception:
+                pass
+        self._reward_workers = None
+
+    # -- Hook overrides ------------------------------------------------------
+
+    def _compute_rewards_for_batch(
+        self, inputs, prompts, completions, completion_ids_list
+    ):
+        """Dispatch rewards to parallel subprocess workers (synchronous wrapper)."""
+        self._launch_reward_workers(inputs, prompts, completions, completion_ids_list)
+        return self._collect_reward_workers(
+            inputs, prompts, completions, completion_ids_list
+        )
+
+    def _launch_reward_workers(self, inputs, prompts, completions, completion_ids_list):
+        """Send reward work to subprocess workers (non-blocking).
+
+        Results are collected later by _collect_reward_workers, allowing GPU
+        logprob computation to overlap with CPU reward computation.
+        """
+        reward_can_bg = all(
+            callable(rf)
+            and not isinstance(rf, nn.Module)
+            and not asyncio.iscoroutinefunction(rf)
+            for rf in self.reward_funcs
+        )
+        num_workers = getattr(self.args, "reward_num_workers", 1)
+
+        if not reward_can_bg or num_workers <= 1:
+            # Can't parallelize — store args for sync fallback in collect
+            self._reward_workers_used = None
+            self._pending_reward_args = (
+                inputs,
+                prompts,
+                completions,
+                completion_ids_list,
+            )
+            return
+
+        workers = self._get_reward_workers()
+        num_generations = self.num_generations
+        num_prompts = len(prompts)
+        num_groups = num_prompts // num_generations
+
+        # Shard by prompt groups across workers
+        groups_per_worker = max(1, (num_groups + len(workers) - 1) // len(workers))
+        workers_used = []
+        for w_idx, (conn, _proc) in enumerate(workers):
+            g_start = w_idx * groups_per_worker
+            g_end = min((w_idx + 1) * groups_per_worker, num_groups)
+            if g_start >= num_groups:
+                break
+            s_start = g_start * num_generations
+            s_end = g_end * num_generations
+            conn.send(
+                (
+                    self.reward_funcs,
+                    prompts[s_start:s_end],
+                    completions[s_start:s_end],
+                    completion_ids_list[s_start:s_end],
+                    inputs[s_start:s_end],
+                    self.reward_func_names,
+                )
+            )
+            workers_used.append(conn)
+
+        self._reward_workers_used = workers_used
+        self._pending_reward_args = (inputs, prompts, completions, completion_ids_list)
+
+    def _collect_reward_workers(
+        self, inputs, prompts, completions, completion_ids_list
+    ):
+        """Collect reward results from subprocess workers (blocks until done)."""
+        from accelerate.utils import gather
+
+        workers_used = getattr(self, "_reward_workers_used", None)
+        args = getattr(self, "_pending_reward_args", None)
+        self._reward_workers_used = None
+        self._pending_reward_args = None
+
+        if workers_used is None:
+            # Sync fallback — compute on main thread
+            if args is not None:
+                return self._calculate_rewards(*args)
+            return self._calculate_rewards(
+                inputs, prompts, completions, completion_ids_list
+            )
+
+        device = self.accelerator.device
+        num_prompts = len(args[1]) if args else len(prompts)
+
+        # Collect results from workers
+        all_worker_results = []
+        any_failed = False
+        for conn in workers_used:
+            result = conn.recv()
+            if result is None:
+                any_failed = True
+                # Drain remaining workers to prevent stale results in pipes
+                for remaining_conn in workers_used:
+                    if remaining_conn is not conn:
+                        try:
+                            remaining_conn.recv()
+                        except Exception:
+                            pass
+                break
+            all_worker_results.append(result)
+
+        if not any_failed:
+            rewards_per_func = torch.zeros(
+                num_prompts, len(self.reward_funcs), device=device
+            )
+            offset = 0
+            for worker_result in all_worker_results:
+                chunk_size = len(worker_result[0])
+                for i, result in enumerate(worker_result):
+                    rewards_per_func[offset : offset + chunk_size, i] = torch.tensor(
+                        result, dtype=torch.float32, device=device
+                    )
+                offset += chunk_size
+            return gather(rewards_per_func)
+
+        # Fallback to main thread on failure
+        if args is not None:
+            return self._calculate_rewards(*args)
+        return self._calculate_rewards(
+            inputs, prompts, completions, completion_ids_list
+        )
+
+    def _post_advantage_hook(
+        self,
+        data: dict,
+        rewards_per_func,
+        advantages,
+        inputs: list,
+        num_generations: int,
+        mode: str,
+        s_start: int | None = None,
+        s_end: int | None = None,
+        is_last_chunk: bool = True,
+    ) -> None:
+        """Replay buffer store/replace + re-roll buffering."""
+        from trl.models.utils import disable_gradient_checkpointing
+
+        # -- Replay buffer: store high-signal groups --
+        if self._replay_buffer is not None:
+            local_grouped = rewards_per_func.view(
+                -1, num_generations, len(self.reward_funcs)
+            )
+            per_group_std = local_grouped.std(dim=1)
+            has_signal = (per_group_std > 0).any(dim=1)
+            offset = s_start or 0
+
+            if has_signal.any():
+                grouped_adv = advantages.view(-1, num_generations)
+                replay_scores = grouped_adv.abs().sum(dim=1) * per_group_std.sum(dim=1)
+                for group_idx in has_signal.nonzero(as_tuple=True)[0]:
+                    gi = group_idx.item()
+                    start = offset + gi * num_generations
+                    end = start + num_generations
+                    group_data = {}
+                    for key in data:
+                        val = data[key]
+                        if (
+                            isinstance(val, torch.Tensor)
+                            and val.dim() > 0
+                            and val.size(0) >= end
+                        ):
+                            group_data[key] = val[start:end].clone()
+                    self._replay_buffer.add(replay_scores[gi].item(), group_data)
+
+            # Replace zero-signal groups with high-signal replay buffer entries
+            # Only in non-streaming path (s_start is None) — streaming scores
+            # groups incrementally, so replacement + logprob recompute would be
+            # too expensive per chunk.
+            n_replaced = 0
+            if s_start is None:
+                no_signal = ~has_signal
+                replaced_ranges = []
+                if no_signal.any() and len(self._replay_buffer) > 0:
+                    for group_idx in no_signal.nonzero(as_tuple=True)[0]:
+                        sampled = self._replay_buffer.sample(1)
+                        if sampled is None:
+                            break
+                        sampled_group = sampled[0]
+                        gi = group_idx.item()
+                        start = offset + gi * num_generations
+                        end = start + num_generations
+                        for key, val in sampled_group.items():
+                            if key in data and isinstance(data[key], torch.Tensor):
+                                src = val.to(data[key].device)
+                                tgt_seq_len = (
+                                    data[key].size(1) if data[key].dim() > 1 else None
+                                )
+                                if start >= data[key].size(0) or end > data[key].size(
+                                    0
+                                ):
+                                    continue
+                                if tgt_seq_len is not None:
+                                    if src.size(1) <= tgt_seq_len:
+                                        data[key][start:end] = 0
+                                        data[key][start:end, : src.size(1)] = src
+                                    else:
+                                        data[key][start:end] = src[:, :tgt_seq_len]
+                                else:
+                                    data[key][start:end] = src
+                        replaced_ranges.append((start, end))
+                        n_replaced += 1
+
+                # Recompute old_per_token_logps for replayed groups
+                if (
+                    n_replaced > 0
+                    and self._replay_recompute_logps
+                    and "old_per_token_logps" in data
+                ):
+                    with (
+                        torch.no_grad(),
+                        disable_gradient_checkpointing(
+                            self.model, self.args.gradient_checkpointing_kwargs
+                        ),
+                    ):
+                        for r_start, r_end in replaced_ranges:
+                            r_ids = torch.cat(
+                                [
+                                    data["prompt_ids"][r_start:r_end],
+                                    data["completion_ids"][r_start:r_end],
+                                ],
+                                dim=1,
+                            )
+                            r_mask = torch.cat(
+                                [
+                                    data["prompt_mask"][r_start:r_end],
+                                    data["completion_mask"][r_start:r_end],
+                                ],
+                                dim=1,
+                            )
+                            r_logits_to_keep = data["completion_ids"].size(1)
+                            r_fwd_kwargs = {}
+                            for fk in (
+                                "pixel_values",
+                                "image_grid_thw",
+                                "pixel_attention_mask",
+                                "image_sizes",
+                                "token_type_ids",
+                                "mm_token_type_ids",
+                            ):
+                                if fk in data:
+                                    r_fwd_kwargs[fk] = data[fk]
+                            r_logps, _ = self._get_per_token_logps_and_entropies(
+                                self.model,
+                                r_ids,
+                                r_mask,
+                                r_logits_to_keep,
+                                r_end - r_start,
+                                **r_fwd_kwargs,
+                            )
+                            data["old_per_token_logps"][r_start:r_end] = r_logps
+
+                if n_replaced > 0:
+                    self._metrics[mode]["replay_buffer_replacements"].append(
+                        float(n_replaced)
+                    )
+
+            if is_last_chunk:
+                self._metrics[mode]["replay_buffer_size"].append(
+                    float(len(self._replay_buffer))
+                )
+
+        # -- Re-roll buffer: store failed prompts --
+        if getattr(self.args, "reroll_start_fraction", 1.0) < 1.0:
+            grouped_rewards = rewards_per_func.view(
+                -1, num_generations, len(self.reward_funcs)
+            )
+            per_group_std = grouped_rewards.std(dim=1)
+            per_group_mean = grouped_rewards.mean(dim=1)
+            zero_signal = (per_group_std == 0).all(dim=1)
+            all_failed = (per_group_mean.abs() < 1e-6).all(dim=1)
+            should_reroll = zero_signal & all_failed
+            _n_buffered = 0
+            with self._reroll_lock:
+                for group_idx in should_reroll.nonzero(as_tuple=True)[0]:
+                    idx = group_idx.item() * num_generations
+                    if idx >= len(inputs):
+                        continue
+                    prompt_input = inputs[idx]
+                    self._reroll_buffer.append(prompt_input)
+                    _n_buffered += 1
+            if _n_buffered > 0:
+                self._metrics[mode]["reroll_buffered"].append(float(_n_buffered))
+            if is_last_chunk:
+                self._metrics[mode]["reroll_buffer_size"].append(
+                    float(len(self._reroll_buffer))
+                )
+
+    # -- Zero-advantage skipping + Liger OPSM ---------------------------------
+
+    def compute_liger_loss(self, unwrapped_model, inputs):
+        """Liger loss with zero-adv skipping and off-policy sequence masking (OPSM).
+
+        The base class Liger path doesn't support OPSM because the fused kernel
+        doesn't expose per-token logprobs needed for the KL computation. This
+        override computes them via chunked lm_head matmul (no grad, low memory)
+        and applies the OPSM to the loss mask before calling the kernel.
+        """
+        if self.args.skip_zero_advantage_batches and torch.all(
+            inputs["advantages"] == 0
+        ):
+            mode = "train" if self.model.training else "eval"
+            self._metrics[mode]["skipped_zero_adv_batches"].append(1.0)
+            return torch.tensor(
+                0.0, device=inputs["advantages"].device, requires_grad=True
+            )
+
+        if self.off_policy_mask_threshold is None:
+            return super().compute_liger_loss(unwrapped_model, inputs)
+
+        # OPSM path: need per_token_logps for KL, which Liger kernel doesn't provide
+        prompt_ids, prompt_mask = inputs["prompt_ids"], inputs["prompt_mask"]
+        completion_ids, completion_mask = (
+            inputs["completion_ids"],
+            inputs["completion_mask"],
+        )
+        input_ids = torch.cat([prompt_ids, completion_ids], dim=1)
+        attention_mask = torch.cat([prompt_mask, completion_mask], dim=1)
+        logits_to_keep = completion_ids.size(1)
+
+        last_hidden_state = self._get_last_hidden_state(
+            unwrapped_model,
+            input_ids,
+            attention_mask,
+            logits_to_keep,
+            inputs.get("pixel_values"),
+            inputs.get("image_grid_thw"),
+            inputs.get("pixel_attention_mask"),
+            inputs.get("image_sizes"),
+        )
+
+        loss_mask = (
+            completion_mask
+            if "tool_mask" not in inputs
+            else completion_mask * inputs["tool_mask"]
+        )
+
+        # Compute per_token_logps via chunked lm_head matmul (no grad, low memory)
+        lm_weight = unwrapped_model.lm_head.weight
+        lm_bias = unwrapped_model.lm_head.bias
+        with torch.no_grad():
+            per_token_logps_chunks = []
+            for i in range(last_hidden_state.size(0)):
+                chunk_logits = torch.matmul(last_hidden_state[i : i + 1], lm_weight.t())
+                if lm_bias is not None:
+                    chunk_logits = chunk_logits + lm_bias
+                chunk_lps = (
+                    chunk_logits.float()
+                    .log_softmax(-1)
+                    .gather(-1, completion_ids[i : i + 1].unsqueeze(-1))
+                    .squeeze(-1)
+                )
+                per_token_logps_chunks.append(chunk_lps)
+                del chunk_logits
+            per_token_logps = torch.cat(per_token_logps_chunks, dim=0)
+
+        advantages = inputs["advantages"]
+        if advantages.dim() == 1:
+            advantages_2d = advantages.unsqueeze(1)
+        else:
+            advantages_2d = advantages
+
+        sampling_per_token_logps = inputs.get("sampling_per_token_logps")
+        if sampling_per_token_logps is None:
+            sampling_per_token_logps = inputs.get("old_per_token_logps")
+        if sampling_per_token_logps is None:
+            sampling_per_token_logps = per_token_logps
+
+        off_policy_mask = GRPOTrainer.get_off_policy_mask(
+            advantages=advantages_2d,
+            per_token_logps=per_token_logps,
+            sampling_per_token_logps=sampling_per_token_logps,
+            mask=loss_mask,
+            off_policy_threshold=self.off_policy_mask_threshold,
+        )
+        loss_mask = loss_mask * off_policy_mask
+
+        # Call the Liger fused kernel with OPSM-modified mask
+        loss, metrics = self.liger_grpo_loss(
+            _input=last_hidden_state,
+            lin_weight=unwrapped_model.lm_head.weight,
+            selected_token_ids=completion_ids,
+            attention_mask=loss_mask,
+            advantages=inputs["advantages"],
+            bias=unwrapped_model.lm_head.bias,
+            old_per_token_logps=inputs.get("old_per_token_logps"),
+            ref_per_token_logps=inputs.get("ref_per_token_logps"),
+            vllm_is_ratio=inputs.get("importance_sampling_ratio"),
+        )
+
+        mean_kl = metrics[0] if self.beta != 0.0 else None
+        clip_ratio = metrics[-1]
+
+        mode = "train" if self.model.training else "eval"
+        if self.beta != 0.0:
+            self._metrics[mode]["kl"].append(
+                self.accelerator.gather(mean_kl).mean().item()
+            )
+        self._metrics[mode]["clip_ratio"].append(
+            self.accelerator.gather(clip_ratio).mean().item()
+        )
+        normalizer = (
+            self.current_gradient_accumulation_steps if mode == "train" else 1.0
+        )
+        return loss / normalizer
+
+    def _compute_loss(self, model, inputs):
+        if self.args.skip_zero_advantage_batches and torch.all(
+            inputs["advantages"] == 0
+        ):
+            mode = "train" if self.model.training else "eval"
+            self._metrics[mode]["skipped_zero_adv_batches"].append(1.0)
+            # Create zero loss with grad_fn. DeepSpeed requires grad_fn != None.
+            # With ZeRO-3, parameters are partitioned (shape=[0], requires_grad=False)
+            # so we can't just do `(p * 0).sum()`. Instead, do a tiny forward pass
+            # with a single token to create a proper computation graph.
+            prompt_ids = inputs["prompt_ids"][:1, :1]  # (1, 1)
+            attn = torch.ones_like(prompt_ids)
+            with torch.amp.autocast(device_type="cuda", dtype=torch.bfloat16):
+                out = model(input_ids=prompt_ids, attention_mask=attn)
+            return out.logits.sum() * 0
+        return super()._compute_loss(model, inputs)

src/axolotl/core/trainers/grpo/replay_buffer.py

@@ -0,0 +1,44 @@
+"""Simple replay buffer for storing and sampling high-signal rollout groups."""
+
+import heapq
+
+import torch
+
+
+class ReplayBuffer:
+    """Min-heap replay buffer that keeps the highest-scoring rollout groups.
+    Groups are scored by signal quality (advantage magnitude * reward variance).
+    When sampling, groups are drawn proportional to their scores.
+    """
+
+    def __init__(self, max_size: int):
+        self.max_size = max_size
+        self._heap: list[tuple[float, int, dict]] = []  # min-heap of (score, id, data)
+        self._counter = 0  # unique tiebreaker for heap
+
+    def __len__(self):
+        return len(self._heap)
+
+    def add(self, score: float, data: dict):
+        """Add a group to the buffer. If full, replaces lowest-scoring entry."""
+        if self.max_size <= 0:
+            return
+        self._counter += 1
+        if len(self._heap) < self.max_size:
+            heapq.heappush(self._heap, (score, self._counter, data))
+        elif score > self._heap[0][0]:
+            heapq.heapreplace(self._heap, (score, self._counter, data))
+
+    def sample(self, num_samples: int) -> list[dict] | None:
+        """Sample groups weighted by their scores. Returns None if buffer is empty."""
+        if self.max_size <= 0 or not self._heap:
+            return None
+
+        scores = torch.tensor([item[0] for item in self._heap], dtype=torch.float32)
+        scores = scores.clamp(min=1e-8)  # avoid zero probabilities
+        probs = scores / scores.sum()
+        replacement = num_samples > len(self._heap)
+        indices = torch.multinomial(
+            probs, num_samples, replacement=replacement
+        ).tolist()
+        return [self._heap[i][2] for i in indices]

src/axolotl/core/trainers/grpo/trainer.py

@@ -40,6 +40,7 @@ from trl.trainer.grpo_config import GRPOConfig
 from trl.trainer.grpo_trainer import RewardFunc, nanstd
 from trl.trainer.utils import pad
 
+from axolotl.core.trainers.grpo.fast_async_trainer import FastAsyncGRPOTrainer
 from axolotl.core.trainers.grpo.sampler import SequenceParallelRepeatRandomSampler
 from axolotl.core.trainers.mixins import (
     DistributedParallelMixin,
@@ -66,6 +67,19 @@ class AxolotlGRPOTrainer(
     _tag_names = ["trl", "grpo", "axolotl"]
 
 
+class AxolotlAsyncGRPOTrainer(
+    RngLoaderMixin,
+    SchedulerMixin,
+    OptimizerMixin,
+    OptimizerInitMixin,
+    DistributedParallelMixin,
+    FastAsyncGRPOTrainer,
+):
+    """Extend AsyncGRPOTrainer with axolotl helpers"""
+
+    _tag_names = ["trl", "grpo", "async", "axolotl"]
+
+
 class AxolotlGRPOSequenceParallelTrainer(AxolotlGRPOTrainer):
     """Extend the base GRPOTrainer for sequence parallelism handling"""
 

src/axolotl/core/trainers/mixins/__init__.py

@@ -4,6 +4,7 @@
 
 from .activation_checkpointing import ActivationOffloadingMixin
 from .checkpoints import CheckpointSaveMixin
+from .layer_offloading import LayerOffloadingMixin
 from .distributed_parallel import DistributedParallelMixin
 from .optimizer import OptimizerMixin
 from .packing import PackingMixin

src/axolotl/core/trainers/mixins/checkpoints.py

@@ -19,5 +19,4 @@ class CheckpointSaveMixin(Trainer):
                 f"Trainer does not support saving optimizer and scheduler:  {exc}\n"
                 "Optimizer and scheduler states were not saved - resuming from checkpoints "
                 "for this training run will not be possible.",
-                main_process_only=True,
             )

src/axolotl/core/trainers/mixins/layer_offloading.py

@@ -0,0 +1,304 @@
+"""
+Trainer mixin for layer-wise parameter offloading to CPU.
+
+Offloads frozen (non-trainable) parameters in decoder layers to CPU, then uses
+forward/backward hooks to stream them on/off GPU one layer at a time with CUDA
+stream prefetching. Trainable parameters (e.g. LoRA weights) stay on GPU always.
+
+Forward:  pre-hook loads layer N's frozen params to GPU (prefetches N+1 on
+          transfer stream), post-hook offloads layer N-1's frozen params.
+Backward: same in reverse order.
+"""
+
+import contextlib
+
+import torch
+import torch.nn as nn
+from transformers import Trainer
+
+from axolotl.utils.logging import get_logger
+
+LOG = get_logger(__name__)
+
+
+def _find_decoder_layers(model: nn.Module) -> tuple[nn.ModuleList | None, list[str]]:
+    """Recursively search the model for the decoder layer ModuleList.
+
+    Finds any ModuleList whose children have 'DecoderLayer' in their class name.
+    Handles all common HF architectures including VLM wrappers (e.g. Qwen3.5-MoE
+    where layers are at model.language_model.layers).
+    """
+    # BFS to find the first ModuleList containing decoder layers
+    queue = [model]
+    while queue:
+        m = queue.pop(0)
+        for _name, child in m.named_children():
+            if isinstance(child, nn.ModuleList) and len(child) > 0:
+                first_type = type(child[0]).__name__
+                if "DecoderLayer" in first_type or "TransformerBlock" in first_type:
+                    layer_types = list({type(layer).__name__ for layer in child})
+                    return child, layer_types
+            else:
+                queue.append(child)
+
+    return None, []
+
+
+def _get_frozen_params(layer: nn.Module) -> list[tuple[str, nn.Parameter]]:
+    """Get all non-trainable parameters in a layer."""
+    return [(n, p) for n, p in layer.named_parameters() if not p.requires_grad]
+
+
+class LayerOffloadManager:
+    """Manages offloading frozen decoder layer params to CPU and streaming
+    them back during forward/backward with CUDA stream overlap.
+
+    Only frozen (requires_grad=False) parameters are offloaded.
+    Trainable parameters (LoRA weights, etc.) remain on GPU at all times.
+    """
+
+    def __init__(
+        self,
+        model: nn.Module,
+        num_prefetch: int = 1,
+    ):
+        self.model = model
+        self.num_prefetch = num_prefetch
+        self._hooks: list = []
+        self._device = None
+
+        # Find decoder layers
+        self.layers, layer_types = _find_decoder_layers(model)
+        if self.layers is None:
+            LOG.warning(
+                "LayerOffloadManager: no decoder layers found, offloading disabled"
+            )
+            self.enabled = False
+            return
+
+        self.enabled = True
+        self.n_layers = len(self.layers)
+        LOG.info(
+            f"Layer offloading: found {self.n_layers} layers ({', '.join(layer_types)})"
+        )
+
+        # Determine GPU device
+        for p in model.parameters():
+            if p.device.type == "cuda":
+                self._device = p.device
+                break
+        if self._device is None:
+            LOG.warning("LayerOffloadManager: no CUDA parameters found")
+            self.enabled = False
+            return
+
+        # Transfer stream for async prefetch
+        self._transfer_stream = torch.cuda.Stream(device=self._device)
+
+        # Track which layers have their frozen params on GPU
+        self._on_gpu: set[int] = set(range(self.n_layers))
+
+        # Cache: frozen param references per layer (list of (name, param) tuples)
+        self._frozen_params: list[list[tuple[str, nn.Parameter]]] = [
+            _get_frozen_params(self.layers[i]) for i in range(self.n_layers)
+        ]
+
+        # CPU storage: pinned tensors for each layer's frozen params
+        # Populated on first offload
+        self._cpu_data: list[dict[str, torch.Tensor]] = [
+            {} for _ in range(self.n_layers)
+        ]
+
+        # Offload all layers upfront
+        self._offload_all()
+
+        # Release cached memory blocks back to the driver
+        torch.cuda.empty_cache()
+
+    def _offload_all(self):
+        """Move all frozen params in all decoder layers to CPU."""
+        mem_before = torch.cuda.memory_allocated(self._device)
+        for i in range(self.n_layers):
+            self._offload_layer(i)
+        mem_after = torch.cuda.memory_allocated(self._device)
+        freed = (mem_before - mem_after) / 1e6
+        LOG.info(
+            f"Layer offloading: offloaded frozen params from {self.n_layers} layers, "
+            f"freed {freed:.0f} MB GPU memory"
+        )
+
+    def _offload_layer(self, idx: int):
+        """Move frozen params of layer idx to CPU pinned memory."""
+        if idx not in self._on_gpu:
+            return
+        for name, param in self._frozen_params[idx]:
+            if param.device.type != "cuda":
+                continue
+            # Allocate pinned CPU tensor on first offload
+            if name not in self._cpu_data[idx]:
+                self._cpu_data[idx][name] = torch.empty_like(
+                    param.data, device="cpu", pin_memory=True
+                )
+            cpu_buf = self._cpu_data[idx][name]
+            # Async copy GPU -> CPU (on transfer stream for overlap)
+            cpu_buf.copy_(param.data, non_blocking=True)
+            # Point parameter at a dummy CPU tensor to free GPU memory
+            param.data = cpu_buf
+        self._on_gpu.discard(idx)
+
+    def _load_layer(self, idx: int, stream=None):
+        """Move frozen params of layer idx back to GPU."""
+        if idx in self._on_gpu or idx < 0 or idx >= self.n_layers:
+            return
+        ctx = (
+            torch.cuda.stream(stream)
+            if stream is not None
+            else contextlib.nullcontext()
+        )
+        with ctx:
+            for _name, param in self._frozen_params[idx]:
+                if param.device.type == "cuda":
+                    continue
+                gpu_data = param.data.to(self._device, non_blocking=True)
+                param.data = gpu_data
+        self._on_gpu.add(idx)
+
+    def _prefetch_layer(self, idx: int):
+        """Async prefetch layer idx on the transfer stream."""
+        if idx in self._on_gpu or idx < 0 or idx >= self.n_layers:
+            return
+        self._transfer_stream.wait_stream(torch.cuda.default_stream(self._device))
+        self._load_layer(idx, stream=self._transfer_stream)
+
+    def _wait_transfer(self):
+        """Make default stream wait for any in-flight transfers."""
+        torch.cuda.default_stream(self._device).wait_stream(self._transfer_stream)
+
+    def setup_hooks(self):
+        """Register forward and backward hooks on each decoder layer."""
+        if not self.enabled:
+            return
+
+        for idx in range(self.n_layers):
+            layer = self.layers[idx]
+
+            def make_pre_fwd(i):
+                def hook(module, args):
+                    # Ensure this layer is on GPU
+                    if i not in self._on_gpu:
+                        self._load_layer(i)
+                    self._wait_transfer()
+                    # Prefetch next layer(s)
+                    for offset in range(1, self.num_prefetch + 1):
+                        self._prefetch_layer(i + offset)
+
+                return hook
+
+            def make_post_fwd(i):
+                def hook(module, args, output):
+                    # Offload previous layer (no longer needed in forward)
+                    if i > 0:
+                        self._offload_layer(i - 1)
+                    # Offload last layer after forward
+                    if i == self.n_layers - 1:
+                        self._offload_layer(i)
+
+                return hook
+
+            def make_pre_bwd(i):
+                def hook(module, grad_output):
+                    # Load this layer for backward
+                    if i not in self._on_gpu:
+                        self._load_layer(i)
+                    self._wait_transfer()
+                    # Prefetch previous layer(s)
+                    for offset in range(1, self.num_prefetch + 1):
+                        self._prefetch_layer(i - offset)
+
+                return hook
+
+            def make_post_bwd(i):
+                def hook(module, grad_input, grad_output):
+                    # Offload the layer above
+                    if i < self.n_layers - 1:
+                        self._offload_layer(i + 1)
+                    # Offload first layer after backward
+                    if i == 0:
+                        self._offload_layer(i)
+
+                return hook
+
+            h1 = layer.register_forward_pre_hook(make_pre_fwd(idx))
+            h2 = layer.register_forward_hook(make_post_fwd(idx))
+            h3 = layer.register_full_backward_pre_hook(make_pre_bwd(idx))
+            h4 = layer.register_full_backward_hook(make_post_bwd(idx))
+            self._hooks.extend([h1, h2, h3, h4])
+
+    def remove_hooks(self):
+        """Remove all hooks and restore layers to GPU."""
+        for h in self._hooks:
+            h.remove()
+        self._hooks.clear()
+        if self.enabled:
+            for i in range(self.n_layers):
+                if i not in self._on_gpu:
+                    self._load_layer(i)
+
+    def pre_step(self):
+        """Called before each training step — ensure layers start offloaded."""
+        if not self.enabled:
+            return
+        for i in list(self._on_gpu):
+            self._offload_layer(i)
+        # Prefetch layer 0 for forward
+        self._prefetch_layer(0)
+
+    def post_step(self):
+        """Called after each training step — ensure layers are offloaded."""
+        if not self.enabled:
+            return
+        for i in list(self._on_gpu):
+            self._offload_layer(i)
+        # Prefetch layer 0 for next step
+        self._prefetch_layer(0)
+
+
+class _LayerOffloadContext:
+    """Context manager wrapping pre_step / post_step around a training step."""
+
+    def __init__(self, manager: LayerOffloadManager):
+        self.manager = manager
+
+    def __enter__(self):
+        self.manager.pre_step()
+        return self
+
+    def __exit__(self, *args):
+        self.manager.post_step()
+
+
+class LayerOffloadingMixin(Trainer):
+    """
+    Trainer mixin class for layer-wise parameter offloading to CPU.
+
+    Offloads frozen decoder layer params to CPU at init, then streams them
+    on/off GPU one layer at a time during each training step.
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        if getattr(self.args, "layer_offloading", False):
+            LOG.info("Layer parameter offloading enabled")
+            self._layer_offload_manager = LayerOffloadManager(
+                model=self.model,
+                num_prefetch=1,
+            )
+            self._layer_offload_manager.setup_hooks()
+            self._layer_offload_ctx = _LayerOffloadContext(self._layer_offload_manager)
+        else:
+            self._layer_offload_manager = None
+            self._layer_offload_ctx = contextlib.nullcontext()
+
+    def training_step(self, *args, **kwargs):
+        with self._layer_offload_ctx:
+            return super().training_step(*args, **kwargs)

src/axolotl/core/training_args_base.py

@@ -235,6 +235,13 @@ class AxolotlTrainingMixins:
         metadata={"help": "Use activation offloading with CUDA streams for training."},
     )
 
+    layer_offloading: bool | None = field(
+        default=None,
+        metadata={
+            "help": "Offload model layer parameters to CPU during forward, prefetch back during backward."
+        },
+    )
+
     # multi-modal section
 
     image_size: int | tuple[int, int] | None = field(

src/axolotl/integrations/cut_cross_entropy/README.md

@@ -19,7 +19,7 @@ python scripts/cutcrossentropy_install.py | sh
 
 - If you are installing from pip
 ```bash
-pip3 uninstall -y cut-cross-entropy && pip3 install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@e8ad129"
+pip3 uninstall -y cut-cross-entropy && pip3 install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@63b15e6"
 ```
 
 ## Usage
@@ -73,8 +73,10 @@ plugins:
 - ministral3
 - mistral
 - mistral3
+- mistral4
 - mixtral
 - mllama
+- nemotron_h
 - olmo
 - olmo2
 - olmo3

src/axolotl/integrations/cut_cross_entropy/__init__.py

@@ -35,7 +35,7 @@ LOG = get_logger(__name__)
 
 _CCE_INSTALL_MESSAGE = (
     "Please install Axolotl's fork of cut_cross_entropy with transformers support using "
-    '`pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@e8ad129"`'
+    '`pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@63b15e6"`'
 )
 
 

src/axolotl/integrations/kernels/autotune_callback.py

@@ -0,0 +1,120 @@
+"""Trainer callback for reporting Triton autotune results from scattermoe-lora kernels."""
+
+import logging
+
+import torch
+from transformers import (
+    TrainerCallback,
+    TrainerControl,
+    TrainerState,
+    TrainingArguments,
+)
+
+LOG = logging.getLogger(__name__)
+
+# Give up looking for autotune data after this many training steps.
+_MAX_POLL_STEP = 5
+
+
+def _get_gpu_info() -> dict:
+    """Return basic GPU identification for the current device."""
+    if not torch.cuda.is_available():
+        return {}
+    try:
+        idx = torch.cuda.current_device()
+        props = torch.cuda.get_device_properties(idx)
+        return {
+            "gpu_name": props.name,
+            "gpu_compute_capability": f"{props.major}.{props.minor}",
+            "gpu_memory_bytes": props.total_memory,
+        }
+    except Exception:  # pylint: disable=broad-exception-caught
+        return {}
+
+
+def _get_smem_capacity() -> dict:
+    """Return shared memory capacity from the runtime lora_ops module."""
+    try:
+        from axolotl.integrations.kernels.autotune_collector import (
+            _find_lora_ops_module,
+        )
+
+        lora_ops = _find_lora_ops_module()
+        if lora_ops is None:
+            return {}
+        fn = getattr(lora_ops, "_get_smem_capacity", None)
+        if fn is None:
+            return {}
+        return {"smem_capacity_bytes": fn()}
+    except Exception:  # pylint: disable=broad-exception-caught
+        return {}
+
+
+class AutotuneReportCallback(TrainerCallback):
+    """Reports Triton kernel autotune selections via telemetry.
+
+    Fires **once** after the first training step completes (step 1), at
+    which point the forward and backward passes have both run and the
+    autotuned kernels have populated their caches.  If for some reason
+    the caches are still empty (e.g. the kernel was never invoked), the
+    callback retries on subsequent steps up to ``_MAX_POLL_STEP`` and
+    then stops polling.
+
+    After reporting (or giving up) every subsequent ``on_step_end``
+    call short-circuits on the ``_reported`` flag — zero hot-path cost.
+    """
+
+    def __init__(self):
+        self._reported = False
+
+    # pylint: disable=unused-argument
+    def on_step_end(
+        self,
+        args: TrainingArguments,
+        state: TrainerState,
+        control: TrainerControl,
+        **kwargs,
+    ):
+        if self._reported:
+            return
+
+        # Lazy import — Triton / scattermoe kernels may not be installed.
+        from axolotl.integrations.kernels.autotune_collector import (
+            collect_autotune_configs,
+        )
+
+        configs = collect_autotune_configs()
+
+        if not configs:
+            if state.global_step >= _MAX_POLL_STEP:
+                LOG.debug(
+                    "No autotune data found after %d steps; giving up.",
+                    state.global_step,
+                )
+                self._reported = True
+            return
+
+        self._reported = True
+
+        from axolotl.telemetry.manager import TelemetryManager
+
+        telemetry_manager = TelemetryManager.get_instance()
+        if not telemetry_manager.enabled:
+            return
+
+        properties = {
+            "kernel_count": len(configs),
+            "kernels": configs,
+        }
+        properties.update(_get_gpu_info())
+        properties.update(_get_smem_capacity())
+
+        telemetry_manager.send_event(
+            event_type="scattermoe-autotune",
+            properties=properties,
+        )
+
+        LOG.info(
+            "Reported %d scattermoe kernel autotune config(s) to telemetry.",
+            len(configs),
+        )

src/axolotl/integrations/kernels/autotune_collector.py

@@ -0,0 +1,114 @@
+"""Collect Triton autotune results from scattermoe-lora kernels.
+
+This module reads the ``.cache`` attribute from Triton ``@triton.autotune``
+decorated kernel objects and returns structured dicts describing the selected
+configurations.  It has **no** telemetry dependency — callers decide what to
+do with the data.
+"""
+
+import logging
+import sys
+from types import ModuleType
+from typing import Any
+
+LOG = logging.getLogger(__name__)
+
+# (human-readable name, attribute on the lora_ops module)
+_KERNEL_REGISTRY: list[tuple[str, str]] = [
+    ("scatter2scatter_lora_fwd", "_scatter2scatter_lora"),
+    ("scatter2scatter_lora_dX", "_scatter2scatter_lora_dX"),
+    ("group_bwd_lora", "_group_bwd_lora"),
+    ("group_bwd_lora_fused", "_group_bwd_lora_fused"),
+]
+
+# The autotune key declared on every kernel: key=["M", "N", "K"]
+_KEY_NAMES: list[str] = ["M", "N", "K"]
+
+
+def _parse_key_tuple(key_tuple: tuple) -> dict[str, Any]:
+    """Turn the autotune cache key tuple into a labelled dict.
+
+    Triton builds the cache key from the values of the declared ``key``
+    args (``M``, ``N``, ``K``) followed by dtype signature elements.
+    We label the first three and store the rest under ``_extra``.
+    """
+    result: dict[str, Any] = {}
+    for i, name in enumerate(_KEY_NAMES):
+        if i < len(key_tuple):
+            result[name] = key_tuple[i]
+    if len(key_tuple) > len(_KEY_NAMES):
+        result["_extra"] = [str(v) for v in key_tuple[len(_KEY_NAMES) :]]
+    return result
+
+
+def _find_lora_ops_module() -> ModuleType | None:
+    """Locate the *runtime* ``lora_ops`` module in ``sys.modules``.
+
+    The HF ``kernels`` package loads ``scattermoe_lora`` via
+    ``import_from_path`` which registers it in ``sys.modules`` under a
+    hash-suffixed name (e.g. ``scattermoe_lora_a1b2c3d4``).  A normal
+    import (``from axolotl.integrations.kernels...``) would create a
+    *separate* module instance whose kernel objects have empty
+    ``.cache`` dicts because autotuning ran on the runtime copy.
+
+    We search ``sys.modules`` for any module whose name contains
+    ``lora_ops`` and that has the ``_scatter2scatter_lora`` kernel
+    attribute — that is the runtime copy with populated caches.
+    """
+    for name, module in list(sys.modules.items()):
+        if (
+            module is not None
+            and "lora_ops" in name
+            and hasattr(module, "_scatter2scatter_lora")
+        ):
+            return module
+    return None
+
+
+def collect_autotune_configs() -> list[dict[str, Any]]:
+    """Read autotune caches from the four scattermoe-lora kernels.
+
+    Returns a (possibly empty) list of dicts, each containing:
+
+    * ``kernel`` – human-readable kernel name
+    * ``key``    – dict with the ``M``/``N``/``K`` problem dimensions
+    * ``config`` – dict with the selected tile sizes, ``num_warps``,
+      and ``num_stages``
+
+    Returns ``[]`` if the kernel module cannot be found or if no
+    autotune cache entries exist yet.
+    """
+    lora_ops = _find_lora_ops_module()
+    if lora_ops is None:
+        LOG.debug(
+            "lora_ops module not found in sys.modules; skipping autotune collection"
+        )
+        return []
+
+    results: list[dict[str, Any]] = []
+
+    for friendly_name, attr_name in _KERNEL_REGISTRY:
+        kernel_fn = getattr(lora_ops, attr_name, None)
+        if kernel_fn is None:
+            continue
+
+        cache = getattr(kernel_fn, "cache", None)
+        if not cache:
+            continue
+
+        for key_tuple, config in cache.items():
+            config_dict = dict(config.kwargs)
+            config_dict["num_warps"] = config.num_warps
+            config_dict["num_stages"] = config.num_stages
+            if getattr(config, "num_ctas", None) is not None:
+                config_dict["num_ctas"] = config.num_ctas
+
+            results.append(
+                {
+                    "kernel": friendly_name,
+                    "key": _parse_key_tuple(key_tuple),
+                    "config": config_dict,
+                }
+            )
+
+    return results

src/axolotl/integrations/kernels/constants.py

@@ -15,6 +15,7 @@ SPARSE_MOE_BLOCK = {
     "qwen2_moe": "Qwen2MoeSparseMoeBlock",
     "qwen3_moe": "Qwen3MoeSparseMoeBlock",
     "qwen3_5_moe": "Qwen3_5MoeSparseMoeBlock",
+    "qwen3_5_moe_text": "Qwen3_5MoeSparseMoeBlock",
     "qwen3_next": "Qwen3NextSparseMoeBlock",
     "qwen3_vl_moe": "Qwen3VLMoeTextSparseMoeBlock",
     # qwen3_omni_moe: Thinker (standard) + Talker (shared experts + shared_expert_gate)
@@ -25,6 +26,8 @@ SPARSE_MOE_BLOCK = {
     "olmoe": "OlmoeSparseMoeBlock",
     "mixtral": "MixtralSparseMoeBlock",
     "minimax": "MiniMaxSparseMoeBlock",
+    # softmax -> topk routing (with group-based expert selection)
+    "mistral4": "Mistral4MoE",
     # sigmoid -> topk routing (with group-based expert selection)
     "glm_moe_dsa": "GlmMoeDsaMoE",
     "deepseek_v3": "DeepseekV3MoE",
@@ -56,7 +59,16 @@ def resolve_moe_block_classes(model_type: str):
 
     cls_names = entry if isinstance(entry, list) else [entry]
     module_path = f"transformers.models.{model_type}.modeling_{model_type}"
-    module = importlib.import_module(module_path)
+    try:
+        module = importlib.import_module(module_path)
+    except ModuleNotFoundError:
+        # Text sub-model types (e.g. qwen3_5_moe_text) share the parent module
+        if model_type.endswith("_text"):
+            parent_type = model_type.removesuffix("_text")
+            module_path = f"transformers.models.{parent_type}.modeling_{parent_type}"
+            module = importlib.import_module(module_path)
+        else:
+            raise
 
     classes = []
     for cls_name in cls_names:

src/axolotl/integrations/kernels/libs/scattermoe_lora/kernels/lora_ops.py

@@ -195,6 +195,36 @@ def _estimate_smem_usage(
 _SMEM_SLACK = 10_000
 
 
+def _estimate_register_pressure(
+    num_warps: int,
+    *tile_sizes: tuple[int, int],
+) -> float:
+    """Rough estimate of per-thread register footprint from live tile sizes.
+
+    This is a heuristic, NOT an accurate register count.  Triton uses tensor
+    core MMA fragments that pack multiple elements per register, and can spill
+    to local memory when the hardware limit (255 regs/thread) is exceeded.
+
+    The estimate is used to prune only truly extreme configs that would cause
+    excessive spilling or compilation failures.  The threshold is set high
+    (``_MAX_REGS_SOFT_LIMIT``) because the heuristic overestimates — it
+    doesn't account for MMA fragment packing.  Configs like M=64,N=64,K=64
+    (est ~520) work fine in practice via spilling.
+
+    Returns estimated registers per thread.
+    """
+    # Each thread in a warp holds ~1/32 of the tile elements
+    tile_regs = sum(r * c for r, c in tile_sizes) / 32
+    scalar_overhead = 40
+    return tile_regs + scalar_overhead
+
+
+# Soft limit for register pressure pruning.  Only prune configs with extreme
+# tile products (e.g. M=128,K=256,N=256) that reliably crash on Blackwell.
+# Moderate configs (M=64,N=64,K=64, est ~520) work via register spilling.
+_MAX_REGS_SOFT_LIMIT = 1024
+
+
 # =============================================================================
 # Forward Kernel: scatter2scatter with fused LoRA
 # =============================================================================
@@ -313,12 +343,11 @@ def _compute_expert_block_lora(
         B_blk_ptrs, mask=N_mask[:, None] & R_mask[None, :], other=0.0
     )  # [BLOCK_N, BLOCK_R]
 
-    # Cast xa_acc and b to same dtype for tl.dot (required when input is bf16/fp16)
-    # Both operands must match; cast to float32 (accumulator type) for precision.
-    b_f32 = b.to(tl.float32)
+    # tl.dot requires non-float32 inputs (tensor cores); cast back to input dtype
+    b_inp = b.to(INPUT_DTYPE)
 
     # (X @ A^T) @ B^T: [M, R] @ [R, N] -> [M, N]
-    lora_out = tl.dot(xa_acc, tl.trans(b_f32), allow_tf32=allow_tf32)
+    lora_out = tl.dot(xa_acc.to(INPUT_DTYPE), tl.trans(b_inp), allow_tf32=allow_tf32)
 
     acc += scaling * lora_out
     return acc
@@ -327,28 +356,29 @@ def _compute_expert_block_lora(
 def _scatter2scatter_lora_configs():
     """Generate forward kernel autotune configs.
 
-    Search space includes smaller tile sizes and fewer pipeline stages to
-    support GPUs with limited shared memory (e.g. ~99KB on some GPUs).
+    Search space includes BLOCK_M to allow trading token-tile size for
+    larger BLOCK_K/BLOCK_N tiles.  On GPUs with ~99KB SMEM, BLOCK_M=128
+    forces BLOCK_K=32 and BLOCK_N=32; BLOCK_M=64 allows BLOCK_K=128
+    (4× fewer inner-loop iterations).
 
     Search space:
-      BLOCK_N:    {32, 64, 128, 256}
+      BLOCK_M:    {32, 64, 128}
+      BLOCK_N:    {32, 64}
       BLOCK_K:    {32, 64, 128}
       num_warps:  {4, 8}
       num_stages: {3, 4, 5}
-
-    BLOCK_M is fixed at 128 (module-level constant, not autotuned in the
-    scatter2scatter pattern).
     """
     configs = []
-    for block_n, block_k, warps, stages in product(
-        [32, 64, 128, 256],  # BLOCK_N
+    for block_m, block_n, block_k, warps, stages in product(
+        [32, 64, 128],  # BLOCK_M
+        [32, 64],  # BLOCK_N
         [32, 64, 128],  # BLOCK_K
         [4, 8],  # num_warps
         [3, 4, 5],  # num_stages
     ):
         configs.append(
             triton.Config(
-                {"BLOCK_N": block_n, "BLOCK_K": block_k},
+                {"BLOCK_M": block_m, "BLOCK_N": block_n, "BLOCK_K": block_k},
                 num_stages=stages,
                 num_warps=warps,
             )
@@ -357,7 +387,7 @@ def _scatter2scatter_lora_configs():
 
 
 def _prune_fwd_configs(configs, named_args, **kwargs):
-    """Prune forward configs based on SMEM capacity.
+    """Prune forward configs based on SMEM capacity and register pressure.
 
     The forward kernel inner loop loads three tiles per pipeline stage:
       X[BLOCK_M, BLOCK_K], W[BLOCK_K, BLOCK_N], A[BLOCK_R, BLOCK_K].
@@ -373,23 +403,49 @@ def _prune_fwd_configs(configs, named_args, **kwargs):
 
     scored = []
     for config in configs:
+        block_m = config.kwargs["BLOCK_M"]
         block_n = config.kwargs["BLOCK_N"]
         block_k = config.kwargs["BLOCK_K"]
         # Base: stages * BLOCK_K * (BLOCK_M + BLOCK_N) + BLOCK_M * BLOCK_N
-        smem_base = _estimate_smem_usage(config.num_stages, BLOCK_M, block_n, block_k)
+        smem_base = _estimate_smem_usage(config.num_stages, block_m, block_n, block_k)
         # A tile [BLOCK_R, BLOCK_K] loaded per stage in the inner loop
         smem_lora_loop = config.num_stages * block_r * block_k * 2
         # B tile [BLOCK_N, BLOCK_R] loaded once in epilogue
         smem_lora_epilogue = block_n * block_r * 2
         smem = smem_base + smem_lora_loop + smem_lora_epilogue
+
+        # Register pressure: live tiles are acc[M,N], xa_acc[M,R],
+        # x[M,K], w[K,N], a[R,K], plus epilogue b[N,R]
+        est_regs = _estimate_register_pressure(
+            config.num_warps,
+            (block_m, block_n),  # acc
+            (block_m, block_r),  # xa_acc
+            (block_m, block_k),  # x tile
+            (block_k, block_n),  # w tile
+            (block_r, block_k),  # a tile
+            (block_n, block_r),  # b tile (epilogue)
+        )
+        if est_regs > _MAX_REGS_SOFT_LIMIT:
+            continue
+
         scored.append((smem, config))
 
     pruned = [c for s, c in scored if s <= smem_cap - _SMEM_SLACK]
     if pruned:
         return pruned
-    # All configs exceed SMEM — return the one with smallest estimated usage
-    scored.sort(key=lambda x: x[0])
-    return [scored[0][1]]
+    if scored:
+        # All surviving configs exceed SMEM — return the one with smallest usage
+        scored.sort(key=lambda x: x[0])
+        return [scored[0][1]]
+    # All configs pruned by register pressure — fall back to smallest tiles
+    return [
+        min(
+            configs,
+            key=lambda c: (
+                c.kwargs["BLOCK_M"] * c.kwargs["BLOCK_N"] * c.kwargs["BLOCK_K"]
+            ),
+        )
+    ]
 
 
 @triton.autotune(
@@ -531,6 +587,89 @@ def _scatter2scatter_lora(
     tl.store(Y_blk_ptrs, acc, mask=M_boundary_mask[:, None] & N_mask[None, :])
 
 
+def _scatter2scatter_lora_split(
+    X: torch.Tensor,
+    W: torch.Tensor,
+    sorted_expert_idxs: torch.Tensor,
+    sorted_scattered_idxs: torch.Tensor,
+    k: int,
+    lora_A: torch.Tensor,
+    lora_B: torch.Tensor,
+    scaling: float,
+    b: Optional[torch.Tensor] = None,
+    x_grouped: bool = False,
+    y_grouped: bool = False,
+    out: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    """Split base+LoRA forward: 3 scatter2scatter calls, no fused LoRA kernel.
+
+    Faster for models with few large experts (e.g. Mixtral E=8, I=14336)
+    because the base kernel runs at full speed without LoRA SMEM overhead,
+    and the LoRA matmuls (R=16) are tiny separate passes.
+
+    Y = scatter(X, W) + scaling * scatter(scatter(X, A^T), B^T)
+    """
+    from axolotl.integrations.kernels.libs.scattermoe_lora.kernels.ops import (
+        scatter2scatter,
+    )
+
+    E = W.size(0)
+    R = lora_A.size(0) // E
+    K = W.size(1)
+    N = W.size(2)
+
+    # 1. Base: Y_base = X @ W  (uses base kernel with optimal tile sizes)
+    output = scatter2scatter(
+        X=X,
+        W=W,
+        b=b,
+        sorted_expert_idxs=sorted_expert_idxs,
+        sorted_scattered_idxs=sorted_scattered_idxs,
+        k=k,
+        x_grouped=x_grouped,
+        y_grouped=y_grouped,
+        out=out,
+    )
+
+    # 2. XA = X @ A^T  (tiny: output is [M*k, R])
+    # Reshape A: [R*E, K] → [E, K, R] (expert weights for scatter2scatter)
+    W_A = lora_A.reshape(E, R, K).permute(0, 2, 1).contiguous()
+    XA = scatter2scatter(
+        X=X,
+        W=W_A,
+        sorted_expert_idxs=sorted_expert_idxs,
+        sorted_scattered_idxs=sorted_scattered_idxs,
+        k=k,
+        x_grouped=x_grouped,
+        y_grouped=True,
+    )
+
+    # 3. Y_lora = XA @ B^T  (R is tiny, so this is very fast)
+    # Reshape B: [N, R*E] → [E, R, N]
+    W_B = lora_B.T.reshape(E, R, N).contiguous()
+    Y_lora = scatter2scatter(
+        X=XA,
+        W=W_B,
+        sorted_expert_idxs=sorted_expert_idxs,
+        sorted_scattered_idxs=sorted_scattered_idxs,
+        k=1,
+        x_grouped=True,
+        y_grouped=y_grouped,
+    )
+
+    # 4. Y = Y_base + scaling * Y_lora
+    output.add_(Y_lora, alpha=scaling)
+    return output
+
+
+# Threshold for switching from fused to split LoRA forward.
+# Split wins when per-expert matmul is large (bandwidth-bound LoRA tile
+# loads dominate in the fused kernel's inner loop).
+# Empirically: split wins for E<=32 with K*N > 20M (e.g. Mixtral, Phi-MoE).
+_SPLIT_LORA_FWD_THRESHOLD = 20_000_000  # per-expert K*N
+_SPLIT_LORA_FWD_MAX_EXPERTS = 32
+
+
 def scatter2scatter_lora(
     X: torch.Tensor,
     W: torch.Tensor,
@@ -546,7 +685,13 @@ def scatter2scatter_lora(
     out: Optional[torch.Tensor] = None,
 ) -> torch.Tensor:
     """
-    Fused scatter2scatter with LoRA: Y[i] = X[i] @ W[e] + scaling * (X[i] @ A[e]^T) @ B[e]^T + b[e]
+    Scatter2scatter with LoRA: Y[i] = X[i] @ W[e] + scaling * (X[i] @ A[e]^T) @ B[e]^T + b[e]
+
+    Automatically selects between:
+    - Fused kernel: single Triton kernel with LoRA in the inner loop.
+      Best for many small experts (E>=64, small K*N).
+    - Split dispatch: 3 separate scatter2scatter calls (base + XA + lora).
+      Best for few large experts (E<=32, large K*N like Mixtral).
 
     Args:
         X: Input [M, K] or [M*k, K] if x_grouped
@@ -565,12 +710,30 @@ def scatter2scatter_lora(
     Returns:
         Y: Output [M*k, N]
     """
-    assert sorted_scattered_idxs.size(0) == sorted_expert_idxs.size(0)
-    assert sorted_scattered_idxs.size(0) == X.size(0) * k
-
     E = W.size(0)
     K = W.size(1)
     N = W.size(2)
+
+    # Dispatch: split for few large experts, fused for many small experts
+    if E <= _SPLIT_LORA_FWD_MAX_EXPERTS and K * N >= _SPLIT_LORA_FWD_THRESHOLD:
+        return _scatter2scatter_lora_split(
+            X,
+            W,
+            sorted_expert_idxs,
+            sorted_scattered_idxs,
+            k,
+            lora_A,
+            lora_B,
+            scaling,
+            b,
+            x_grouped,
+            y_grouped,
+            out,
+        )
+
+    assert sorted_scattered_idxs.size(0) == sorted_expert_idxs.size(0)
+    assert sorted_scattered_idxs.size(0) == X.size(0) * k
+
     R = lora_A.size(0) // E
 
     # Pad R to power of 2 for Triton tile size
@@ -610,11 +773,9 @@ def scatter2scatter_lora(
         b_ptr,
         stride_be,
         stride_bn,
-        # A: [r*E, K] -> stride(0) is r*E dim stride, stride(1) is K dim stride
         lora_A,
         lora_A.stride(0),
         lora_A.stride(1),
-        # B: [N, r*E] -> stride(0) is N dim stride, stride(1) is r*E dim stride
         lora_B,
         lora_B.stride(0),
         lora_B.stride(1),
@@ -625,9 +786,8 @@ def scatter2scatter_lora(
         K=K,
         N=N,
         E=E,
-        ACTUAL_R=R,  # True LoRA rank for weight indexing
-        BLOCK_M=BLOCK_M,
-        BLOCK_R=BLOCK_R,  # Padded tile size >= max(R, 16)
+        ACTUAL_R=R,
+        BLOCK_R=BLOCK_R,
         ACC_TYPE=tl.float32,
         scaling=scaling,
         allow_tf32=ALLOW_TF32,
@@ -761,13 +921,13 @@ def _compute_expert_block_lora_dX(
         + (A_expert_offset + R_block)[:, None] * stride_ar
         + K_block[None, :] * stride_ak
     )
-    a_e = tl.load(A_blk_ptrs, mask=R_mask[:, None] & K_mask[None, :], other=0.0)
-
-    # Cast to float32 for precision
-    a_f32 = a_e.to(tl.float32)
+    a_e = tl.load(A_blk_ptrs, mask=R_mask[:, None] & K_mask[None, :], other=0.0).to(
+        INPUT_DTYPE
+    )
 
     # (DY @ B) @ A: [M, R] @ [R, K] -> [M, K]
-    lora_dx = tl.dot(dy_b_acc, a_f32, allow_tf32=allow_tf32)
+    # tl.dot requires non-float32 inputs (tensor cores); cast accumulator back to input dtype
+    lora_dx = tl.dot(dy_b_acc.to(INPUT_DTYPE), a_e, allow_tf32=allow_tf32)
 
     acc += scaling * lora_dx
     return acc
@@ -779,25 +939,26 @@ def _scatter2scatter_lora_dX_configs():
     The inner loop is over N (not K as in forward). The output dimension is K.
     So BLOCK_K tiles the output and BLOCK_N tiles the reduction.
 
-    Search space includes smaller tile sizes and fewer pipeline stages to
-    support GPUs with limited shared memory (e.g. ~99KB on some GPUs).
+    BLOCK_M is now autotunable (was fixed at 128).
 
     Search space:
-      BLOCK_K:    {32, 64, 128, 256}   (output tile)
-      BLOCK_N:    {32, 64, 128, 256}   (reduction tile)
+      BLOCK_M:    {32, 64, 128}        (token tile)
+      BLOCK_K:    {32, 64, 128}   (output tile)
+      BLOCK_N:    {32, 64}   (reduction tile)
       num_warps:  {4, 8}
       num_stages: {3, 4, 5}
     """
     configs = []
-    for block_k, block_n, warps, stages in product(
-        [32, 64, 128, 256],  # BLOCK_K (output dimension)
-        [32, 64, 128, 256],  # BLOCK_N (reduction dimension)
+    for block_m, block_k, block_n, warps, stages in product(
+        [32, 64, 128],  # BLOCK_M
+        [32, 64, 128],  # BLOCK_K (output dimension)
+        [32, 64],  # BLOCK_N (reduction dimension)
         [4, 8],  # num_warps
         [3, 4, 5],  # num_stages
     ):
         configs.append(
             triton.Config(
-                {"BLOCK_K": block_k, "BLOCK_N": block_n},
+                {"BLOCK_M": block_m, "BLOCK_K": block_k, "BLOCK_N": block_n},
                 num_stages=stages,
                 num_warps=warps,
             )
@@ -806,7 +967,7 @@ def _scatter2scatter_lora_dX_configs():
 
 
 def _prune_dX_configs(configs, named_args, **kwargs):
-    """Prune backward dX configs based on SMEM capacity.
+    """Prune backward dX configs based on SMEM capacity and register pressure.
 
     The dX kernel inner loop loads three tiles per pipeline stage:
       DY[BLOCK_M, BLOCK_N], W^T[BLOCK_N, BLOCK_K], B[BLOCK_N, BLOCK_R].
@@ -822,23 +983,49 @@ def _prune_dX_configs(configs, named_args, **kwargs):
 
     scored = []
     for config in configs:
+        block_m = config.kwargs["BLOCK_M"]
         block_k = config.kwargs["BLOCK_K"]
         block_n = config.kwargs["BLOCK_N"]
         # Base: stages * BLOCK_N * (BLOCK_M + BLOCK_K) + BLOCK_M * BLOCK_K
-        smem_base = _estimate_smem_usage(config.num_stages, BLOCK_M, block_k, block_n)
+        smem_base = _estimate_smem_usage(config.num_stages, block_m, block_k, block_n)
         # B tile [BLOCK_N, BLOCK_R] loaded per stage in the inner loop
         smem_lora_loop = config.num_stages * block_n * block_r * 2
         # A tile [BLOCK_R, BLOCK_K] loaded once in epilogue
         smem_lora_epilogue = block_r * block_k * 2
         smem = smem_base + smem_lora_loop + smem_lora_epilogue
+
+        # Register pressure: live tiles are acc[M,K], dy_b_acc[M,R],
+        # dy[M,N], wt[N,K], b[N,R], plus epilogue a[R,K]
+        est_regs = _estimate_register_pressure(
+            config.num_warps,
+            (block_m, block_k),  # acc
+            (block_m, block_r),  # dy_b_acc
+            (block_m, block_n),  # dy tile
+            (block_n, block_k),  # wt tile
+            (block_n, block_r),  # b tile
+            (block_r, block_k),  # a tile (epilogue)
+        )
+        if est_regs > _MAX_REGS_SOFT_LIMIT:
+            continue
+
         scored.append((smem, config))
 
     pruned = [c for s, c in scored if s <= smem_cap - _SMEM_SLACK]
     if pruned:
         return pruned
-    # All configs exceed SMEM — return the one with smallest estimated usage
-    scored.sort(key=lambda x: x[0])
-    return [scored[0][1]]
+    if scored:
+        # All surviving configs exceed SMEM — return the one with smallest usage
+        scored.sort(key=lambda x: x[0])
+        return [scored[0][1]]
+    # All configs pruned by register pressure — fall back to smallest tiles
+    return [
+        min(
+            configs,
+            key=lambda c: (
+                c.kwargs["BLOCK_M"] * c.kwargs["BLOCK_K"] * c.kwargs["BLOCK_N"]
+            ),
+        )
+    ]
 
 
 @triton.autotune(
@@ -1067,7 +1254,7 @@ def scatter2scatter_lora_dX(
         N=N,
         E=E,
         ACTUAL_R=R,
-        BLOCK_M=BLOCK_M,
+        # BLOCK_M is autotuned (injected by triton.autotune from Config kwargs)
         BLOCK_R=BLOCK_R,
         ACC_TYPE=tl.float32,
         scaling=scaling,
@@ -1091,9 +1278,9 @@ def _group_bwd_lora_configs():
     support GPUs with limited shared memory (e.g. ~99KB on some GPUs).
 
     Search space:
-      BLOCK_M:    {32, 64, 128, 256}   (token-loop tile)
-      BLOCK_K:    {32, 64, 128, 256}
-      BLOCK_N:    {32, 64, 128, 256}
+      BLOCK_M:    {32, 64, 128}   (token-loop tile)
+      BLOCK_K:    {32, 64, 128}
+      BLOCK_N:    {32, 64}
       num_warps:  {4, 8}
       num_stages: {3, 4, 5}
 
@@ -1102,9 +1289,9 @@ def _group_bwd_lora_configs():
     """
     configs = []
     for block_m, block_k, block_n, warps, stages in product(
-        [32, 64, 128, 256],  # BLOCK_M
-        [32, 64, 128, 256],  # BLOCK_K
-        [32, 64, 128, 256],  # BLOCK_N
+        [32, 64, 128],  # BLOCK_M
+        [32, 64, 128],  # BLOCK_K
+        [32, 64],  # BLOCK_N
         [4, 8],  # num_warps
         [3, 4, 5],  # num_stages
     ):
@@ -1119,7 +1306,7 @@ def _group_bwd_lora_configs():
 
 
 def _prune_bwd_lora_configs(configs, named_args, **kwargs):
-    """Prune backward configs based on SMEM capacity.
+    """Prune backward configs based on SMEM capacity and register pressure.
 
     The backward kernel loads X[BLOCK_M, BLOCK_K] and DY[BLOCK_M, BLOCK_N]
     in the inner loop, plus holds A[BLOCK_R, BLOCK_K] and B[BLOCK_N, BLOCK_R]
@@ -1138,14 +1325,40 @@ def _prune_bwd_lora_configs(configs, named_args, **kwargs):
         # A[BLOCK_R, BLOCK_K] and B[BLOCK_N, BLOCK_R] held for the full expert
         smem_lora = (block_r * block_k + block_n * block_r) * 2
         smem = smem_base + smem_lora
+
+        # Register pressure: dA_acc[R,K], dB_acc[N,R], x[M,K], dy[M,N],
+        # a[R,K], b[N,R], xa[M,R], dy_b[M,R]
+        est_regs = _estimate_register_pressure(
+            config.num_warps,
+            (block_r, block_k),  # dA_acc
+            (block_n, block_r),  # dB_acc
+            (block_m, block_k),  # x tile
+            (block_m, block_n),  # dy tile
+            (block_r, block_k),  # a tile
+            (block_n, block_r),  # b tile
+            (block_m, block_r),  # xa intermediate
+        )
+        if est_regs > _MAX_REGS_SOFT_LIMIT:
+            continue
+
         scored.append((smem, config))
 
     pruned = [c for s, c in scored if s <= smem_cap - _SMEM_SLACK]
     if pruned:
         return pruned
-    # All configs exceed SMEM — return the one with smallest estimated usage
-    scored.sort(key=lambda x: x[0])
-    return [scored[0][1]]
+    if scored:
+        # All surviving configs exceed SMEM — return the one with smallest usage
+        scored.sort(key=lambda x: x[0])
+        return [scored[0][1]]
+    # All configs pruned by register pressure — fall back to smallest tiles
+    return [
+        min(
+            configs,
+            key=lambda c: (
+                c.kwargs["BLOCK_M"] * c.kwargs["BLOCK_K"] * c.kwargs["BLOCK_N"]
+            ),
+        )
+    ]
 
 
 @triton.autotune(
@@ -1330,6 +1543,279 @@ def _group_bwd_lora(
         )
 
 
+def _group_bwd_split_configs():
+    """Autotune configs for split dA/dB kernels."""
+    configs = []
+    for block_m, block_dim, warps, stages in product(
+        [32, 64, 128],  # BLOCK_M (token tile)
+        [32, 64, 128, 256],  # BLOCK_DIM (K for dA, N for dB — output tile)
+        [4, 8],  # num_warps
+        [3, 4, 5],  # num_stages
+    ):
+        configs.append(
+            triton.Config(
+                {"BLOCK_M": block_m, "BLOCK_DIM": block_dim},
+                num_stages=stages,
+                num_warps=warps,
+            )
+        )
+    return configs
+
+
+def _prune_split_configs(configs, named_args, **kwargs):
+    """Prune split kernel configs based on SMEM capacity and register pressure."""
+    smem_cap = _get_smem_capacity()
+    block_r = named_args.get("BLOCK_R", 64)
+
+    # Fixed inner tile for reduction dimension
+    BLOCK_INNER = 64
+
+    pruned = []
+    for config in configs:
+        block_m = config.kwargs["BLOCK_M"]
+        block_dim = config.kwargs["BLOCK_DIM"]
+        # Inner loop loads: input[M, INNER] and other[M, INNER_or_DIM]
+        smem = config.num_stages * BLOCK_INNER * (block_m + block_dim) * 2
+        # LoRA weights held in registers: [INNER, R] or [R, DIM]
+        smem += (block_r * max(block_dim, BLOCK_INNER)) * 2
+
+        # Register pressure check
+        est_regs = _estimate_register_pressure(
+            config.num_warps,
+            (block_r, block_dim),  # acc
+            (block_m, BLOCK_INNER),  # input tile
+            (block_m, block_dim),  # other tile
+            (block_r, BLOCK_INNER),  # lora weight
+        )
+        if est_regs > _MAX_REGS_SOFT_LIMIT:
+            continue
+
+        if smem <= smem_cap - _SMEM_SLACK:
+            pruned.append(config)
+
+    if pruned:
+        return pruned
+    configs.sort(key=lambda c: c.kwargs["BLOCK_M"] * c.kwargs["BLOCK_DIM"])
+    return [configs[0]]
+
+
+@triton.autotune(
+    configs=_group_bwd_split_configs(),
+    key=["M", "K", "N"],
+    prune_configs_by={"early_config_prune": _prune_split_configs},
+)
+@triton.heuristics(
+    {
+        "NO_DIM_MASK": lambda args: (
+            (args["K"] % args["BLOCK_DIM"]) == 0
+            if args["COMPUTE_DA"]
+            else (args["N"] % args["BLOCK_DIM"]) == 0
+        ),
+    }
+)
+@triton.jit
+def _group_bwd_lora_split(
+    # Data tensors (DY and X are always present)
+    DY_ptr,
+    stride_dym,
+    stride_dyn,
+    X_ptr,
+    stride_xm,
+    stride_xk,
+    # LoRA weight for the inner reduction (B for dA, A for dB)
+    LW_ptr,
+    stride_lw0,
+    stride_lw1,
+    # Output gradient tensor (dA or dB)
+    OUT_ptr,
+    stride_out0,
+    stride_out1,
+    # Expert offsets
+    expert_offsets_ptr,
+    # Dimensions
+    M,
+    K: tl.constexpr,
+    N: tl.constexpr,
+    ACTUAL_R: tl.constexpr,
+    BLOCK_R: tl.constexpr,
+    INNER_DIM: tl.constexpr,  # reduction dimension (N for dA, K for dB)
+    scaling,
+    # Mode flag
+    COMPUTE_DA: tl.constexpr,  # True = compute dA, False = compute dB
+    # Tile sizes
+    BLOCK_M: tl.constexpr,
+    BLOCK_DIM: tl.constexpr,
+    ACC_TYPE: tl.constexpr,
+    allow_tf32: tl.constexpr,
+    NO_DIM_MASK: tl.constexpr,
+):
+    """
+    Unified split kernel for LoRA gradient computation.
+
+    When COMPUTE_DA=True:
+      dA[e] = scaling * (dY @ B[e])^T @ X  →  [R, K]
+      Grid: (E, cdiv(K, BLOCK_DIM))
+      - outer_ptr/stride = X (read [M, K_block])
+      - inner reduction over N using DY and B
+      - output shape [BLOCK_R, BLOCK_DIM]
+
+    When COMPUTE_DA=False:
+      dB[e] = scaling * dY^T @ (X @ A[e]^T)  →  [N, R]
+      Grid: (E, cdiv(N, BLOCK_DIM))
+      - outer_ptr/stride = DY (read [M, N_block])
+      - inner reduction over K using X and A
+      - output shape [BLOCK_DIM, BLOCK_R]
+
+    No atomic adds — each (E, dim_block) pair is written by exactly one block.
+    """
+    E_idx = tl.program_id(0)
+    dim_block_id = tl.program_id(1)
+
+    if E_idx == 0:
+        start_idx = 0
+    else:
+        start_idx = tl.load(expert_offsets_ptr + E_idx - 1).to(tl.int32)
+    end_idx = tl.load(expert_offsets_ptr + E_idx).to(tl.int32)
+    num_tokens = end_idx - start_idx
+
+    # Output dimension tile (K for dA, N for dB)
+    if COMPUTE_DA:
+        OUT_DIM: tl.constexpr = K  # type: ignore[no-redef]
+    else:
+        OUT_DIM: tl.constexpr = N  # type: ignore[no-redef]
+    dim_block = dim_block_id * BLOCK_DIM + tl.arange(0, BLOCK_DIM)
+    dim_mask = dim_block < OUT_DIM
+    R_block = tl.arange(0, BLOCK_R)
+    R_mask = R_block < ACTUAL_R
+    lora_offset = E_idx * ACTUAL_R
+
+    # Output pointers — layout differs: dA is [R, K], dB is [N, R]
+    if COMPUTE_DA:
+        out_blk_ptrs = (
+            OUT_ptr
+            + (lora_offset + R_block)[:, None] * stride_out0
+            + dim_block[None, :] * stride_out1
+        )
+        out_mask = R_mask[:, None] & dim_mask[None, :]
+    else:
+        out_blk_ptrs = (
+            OUT_ptr
+            + dim_block[:, None] * stride_out0
+            + (lora_offset + R_block)[None, :] * stride_out1
+        )
+        out_mask = dim_mask[:, None] & R_mask[None, :]
+
+    if num_tokens > 0:
+        M_block = tl.arange(0, BLOCK_M)
+        INPUT_DTYPE = X_ptr.dtype.element_ty
+        BLOCK_INNER: tl.constexpr = 64
+        inner_iters = tl.cdiv(INNER_DIM, BLOCK_INNER)
+
+        if COMPUTE_DA:
+            acc = tl.zeros((BLOCK_R, BLOCK_DIM), dtype=ACC_TYPE)
+        else:
+            acc = tl.zeros((BLOCK_DIM, BLOCK_R), dtype=ACC_TYPE)
+
+        M_iters = tl.cdiv(num_tokens, BLOCK_M)
+        for i in range(M_iters):
+            M_idx = start_idx + i * BLOCK_M + M_block
+            M_mask = M_idx < end_idx
+
+            if COMPUTE_DA:
+                # Load X[M, K_block] (the "outer" tensor for dA)
+                outer = tl.load(
+                    X_ptr + M_idx[:, None] * stride_xm + dim_block[None, :] * stride_xk,
+                    mask=M_mask[:, None] & dim_mask[None, :],
+                    other=0.0,
+                ).to(INPUT_DTYPE)
+
+                # Reduce DY[M, :] @ B[e][:, R] over N → [M, R]
+                reduced = tl.zeros((BLOCK_M, BLOCK_R), dtype=ACC_TYPE)
+                inner_range = tl.arange(0, BLOCK_INNER)
+                for j in range(inner_iters):
+                    inn_off = j * BLOCK_INNER + inner_range
+                    inn_mask = inn_off < N
+
+                    dy_tile = tl.load(
+                        DY_ptr
+                        + M_idx[:, None] * stride_dym
+                        + inn_off[None, :] * stride_dyn,
+                        mask=M_mask[:, None] & inn_mask[None, :],
+                        other=0.0,
+                    ).to(INPUT_DTYPE)
+                    # B layout: [N, r*E] → stride_lw0=N stride, stride_lw1=r*E stride
+                    lw_tile = tl.load(
+                        LW_ptr
+                        + inn_off[:, None] * stride_lw0
+                        + (lora_offset + R_block)[None, :] * stride_lw1,
+                        mask=inn_mask[:, None] & R_mask[None, :],
+                        other=0.0,
+                    ).to(INPUT_DTYPE)
+                    reduced += tl.dot(dy_tile, lw_tile, allow_tf32=allow_tf32)
+
+                # dA += (DY@B)^T @ X: [R, M] @ [M, K_block] → [R, K_block]
+                acc += tl.dot(
+                    tl.trans(reduced.to(INPUT_DTYPE)), outer, allow_tf32=allow_tf32
+                )
+            else:
+                # Load DY[M, N_block] (the "outer" tensor for dB)
+                outer = tl.load(
+                    DY_ptr
+                    + M_idx[:, None] * stride_dym
+                    + dim_block[None, :] * stride_dyn,
+                    mask=M_mask[:, None] & dim_mask[None, :],
+                    other=0.0,
+                ).to(INPUT_DTYPE)
+
+                # Reduce X[M, :] @ A[e][:, :].T over K → [M, R]
+                reduced = tl.zeros((BLOCK_M, BLOCK_R), dtype=ACC_TYPE)
+                inner_range = tl.arange(0, BLOCK_INNER)
+                for j in range(inner_iters):
+                    inn_off = j * BLOCK_INNER + inner_range
+                    inn_mask = inn_off < K
+
+                    x_tile = tl.load(
+                        X_ptr
+                        + M_idx[:, None] * stride_xm
+                        + inn_off[None, :] * stride_xk,
+                        mask=M_mask[:, None] & inn_mask[None, :],
+                        other=0.0,
+                    ).to(INPUT_DTYPE)
+                    # A layout: [r*E, K] → stride_lw0=r*E stride, stride_lw1=K stride
+                    # We want A[e]^T: [K, R], so load as [K_inner, R]
+                    lw_tile = tl.load(
+                        LW_ptr
+                        + (lora_offset + R_block)[None, :] * stride_lw0
+                        + inn_off[:, None] * stride_lw1,
+                        mask=inn_mask[:, None] & R_mask[None, :],
+                        other=0.0,
+                    ).to(INPUT_DTYPE)
+                    reduced += tl.dot(x_tile, lw_tile, allow_tf32=allow_tf32)
+
+                # dB += DY^T @ (X@A^T): [N_block, M] @ [M, R] → [N_block, R]
+                acc += tl.dot(
+                    tl.trans(outer), reduced.to(INPUT_DTYPE), allow_tf32=allow_tf32
+                )
+
+        tl.store(
+            out_blk_ptrs, (acc * scaling).to(OUT_ptr.dtype.element_ty), mask=out_mask
+        )
+    else:
+        # Zero out this expert's slice — needed because output uses empty_like
+        if COMPUTE_DA:
+            tl.store(
+                out_blk_ptrs,
+                tl.zeros((BLOCK_R, BLOCK_DIM), dtype=OUT_ptr.dtype.element_ty),
+                mask=out_mask,
+            )
+        else:
+            tl.store(
+                out_blk_ptrs,
+                tl.zeros((BLOCK_DIM, BLOCK_R), dtype=OUT_ptr.dtype.element_ty),
+                mask=out_mask,
+            )
+
+
 def group_bwd_lora(
     DY: torch.Tensor,
     X: torch.Tensor,
@@ -1344,6 +1830,9 @@ def group_bwd_lora(
     """
     Compute LoRA gradients for A and B on expert-grouped data.
 
+    Uses split dA/dB kernels that eliminate atomic adds by giving each
+    (expert, output_block) pair its own thread block.
+
     Args:
         DY: Gradient w.r.t. output [M_total, N] (grouped by expert)
         X: Input [M_total, K] (grouped by expert)
@@ -1361,19 +1850,46 @@ def group_bwd_lora(
     K = X.size(1)
     N = DY.size(1)
 
-    # Zero-init for atomic accumulation
-    dA = torch.zeros_like(lora_A)
-    dB = torch.zeros_like(lora_B)
+    # No zero-init needed: the split kernels write zeros for experts with
+    # zero routed tokens directly in the kernel (else branch).
+    dA = torch.empty_like(lora_A)
+    dB = torch.empty_like(lora_B)
 
     BLOCK_R = _block_r_for_rank(R)
 
-    def grid(META):
-        return (
-            E * triton.cdiv(K, META["BLOCK_K"]),
-            triton.cdiv(N, META["BLOCK_N"]),
-        )
+    def grid_dA(META):
+        return (E, triton.cdiv(K, META["BLOCK_DIM"]))
 
-    _group_bwd_lora[grid](
+    _group_bwd_lora_split[grid_dA](
+        DY,
+        DY.stride(0),
+        DY.stride(1),
+        X,
+        X.stride(0),
+        X.stride(1),
+        lora_B,
+        lora_B.stride(0),
+        lora_B.stride(1),
+        dA,
+        dA.stride(0),
+        dA.stride(1),
+        expert_offsets,
+        M=DY.size(0),
+        K=K,
+        N=N,
+        ACTUAL_R=R,
+        BLOCK_R=BLOCK_R,
+        INNER_DIM=N,
+        scaling=scaling,
+        COMPUTE_DA=True,
+        ACC_TYPE=tl.float32,
+        allow_tf32=ALLOW_TF32,
+    )
+
+    def grid_dB(META):
+        return (E, triton.cdiv(N, META["BLOCK_DIM"]))
+
+    _group_bwd_lora_split[grid_dB](
         DY,
         DY.stride(0),
         DY.stride(1),
@@ -1383,12 +1899,6 @@ def group_bwd_lora(
         lora_A,
         lora_A.stride(0),
         lora_A.stride(1),
-        lora_B,
-        lora_B.stride(0),
-        lora_B.stride(1),
-        dA,
-        dA.stride(0),
-        dA.stride(1),
         dB,
         dB.stride(0),
         dB.stride(1),
@@ -1396,9 +1906,11 @@ def group_bwd_lora(
         M=DY.size(0),
         K=K,
         N=N,
-        ACTUAL_R=R,  # True LoRA rank
-        BLOCK_R=BLOCK_R,  # Padded tile size
+        ACTUAL_R=R,
+        BLOCK_R=BLOCK_R,
+        INNER_DIM=K,
         scaling=scaling,
+        COMPUTE_DA=False,
         ACC_TYPE=tl.float32,
         allow_tf32=ALLOW_TF32,
     )

src/axolotl/integrations/kernels/libs/scattermoe_lora/layers.py

@@ -220,6 +220,158 @@ def _unwrap_experts_lora(experts_module):
     return base_experts, gup_lora, down_lora
 
 
+# =============================================================================
+# Routing helpers
+# =============================================================================
+
+
+def _softmax_topk_route(
+    moe_block, base_gate, hidden_states, gate_weight, gate_lora_delta
+):
+    """Softmax→topk routing (Qwen, OLMoE, Mixtral, MiniMax).
+
+    Returns:
+        (routing_weights [T, K], selected_experts [T, K], top_k, num_experts)
+    """
+    router_logits = F.linear(hidden_states, gate_weight)
+    if gate_lora_delta is not None:
+        router_logits = router_logits + F.linear(hidden_states, gate_lora_delta)
+    routing_weights = F.softmax(router_logits, dim=-1, dtype=torch.float32)
+
+    top_k = base_gate.top_k
+    num_experts = base_gate.num_experts
+    routing_weights, selected_experts = torch.topk(routing_weights, top_k, dim=-1)
+
+    if getattr(base_gate, "norm_topk_prob", True):
+        routing_weights = routing_weights / routing_weights.sum(dim=-1, keepdim=True)
+
+    return routing_weights, selected_experts, top_k, num_experts
+
+
+def _sigmoid_topk_route(
+    moe_block, base_gate, hidden_states, gate_weight, gate_lora_delta
+):
+    """Sigmoid→topk routing (GLM, DeepSeek V3, MiniMax M2).
+
+    Supports:
+    - ``e_score_correction_bias`` on gate or moe_block
+    - Group-based expert selection when ``n_group > 1``
+    - ``routed_scaling_factor`` applied to final weights
+    - Final weights gathered from original sigmoid probs (not bias-corrected)
+
+    Returns:
+        (routing_weights [T, K], selected_experts [T, K], top_k, num_experts)
+    """
+    router_logits = F.linear(hidden_states.float(), gate_weight.float())
+    if gate_lora_delta is not None:
+        router_logits = router_logits + F.linear(
+            hidden_states.float(), gate_lora_delta.float()
+        )
+    router_probs = router_logits.sigmoid()  # [T, E]
+
+    top_k = getattr(moe_block, "top_k", getattr(base_gate, "top_k", None))
+    num_experts = getattr(moe_block, "n_routed_experts", gate_weight.shape[0])
+
+    # Bias-corrected scores for expert selection (not used for final weights).
+    # glm_moe_dsa/deepseek_v3 store the bias on gate; minimax_m2 on the block.
+    e_score_correction_bias = getattr(base_gate, "e_score_correction_bias", None)
+    if e_score_correction_bias is None:
+        e_score_correction_bias = getattr(moe_block, "e_score_correction_bias", None)
+    if e_score_correction_bias is not None:
+        scores_for_choice = router_probs + e_score_correction_bias
+    else:
+        scores_for_choice = router_probs
+
+    # Group-based selection: pick top groups, mask the rest
+    n_group = getattr(moe_block, "n_group", 1)
+    if n_group > 1:
+        group_scores = (
+            scores_for_choice.view(-1, n_group, num_experts // n_group)
+            .topk(2, dim=-1)[0]
+            .sum(dim=-1)
+        )  # [T, n_group]
+        topk_group = getattr(moe_block, "topk_group", n_group)
+        group_idx = torch.topk(group_scores, k=topk_group, dim=-1, sorted=False)[1]
+        group_mask = torch.zeros_like(group_scores)
+        group_mask.scatter_(1, group_idx, 1)
+        score_mask = (
+            group_mask.unsqueeze(-1)
+            .expand(-1, n_group, num_experts // n_group)
+            .reshape(-1, num_experts)
+        )
+        scores_for_choice = scores_for_choice.masked_fill(~score_mask.bool(), 0.0)
+
+    # Final topk from (possibly masked) scores
+    topk_indices = torch.topk(scores_for_choice, k=top_k, dim=-1, sorted=False)[1]
+
+    # Gather weights from original sigmoid scores (not bias-corrected)
+    topk_weights = router_probs.gather(1, topk_indices)
+
+    # Optional renormalization + scaling
+    if getattr(moe_block, "norm_topk_prob", True):
+        topk_weights = topk_weights / (topk_weights.sum(dim=-1, keepdim=True) + 1e-20)
+    routed_scaling_factor = getattr(moe_block, "routed_scaling_factor", 1.0)
+    topk_weights = topk_weights * routed_scaling_factor
+
+    return topk_weights, topk_indices, top_k, num_experts
+
+
+def _route(moe_block, base_gate, hidden_states, gate_weight, gate_lora_delta):
+    """Dispatch to the correct routing strategy based on block attributes.
+
+    Detects sigmoid routing by the presence of ``e_score_correction_bias``
+    on either the gate or the moe_block.
+    """
+    has_sigmoid = (
+        getattr(base_gate, "e_score_correction_bias", None) is not None
+        or getattr(moe_block, "e_score_correction_bias", None) is not None
+    )
+    if has_sigmoid:
+        return _sigmoid_topk_route(
+            moe_block, base_gate, hidden_states, gate_weight, gate_lora_delta
+        )
+    return _softmax_topk_route(
+        moe_block, base_gate, hidden_states, gate_weight, gate_lora_delta
+    )
+
+
+# =============================================================================
+# Shared expert helpers
+# =============================================================================
+
+
+def _compute_shared_expert(moe_block, hidden_states_flat):
+    """Compute shared expert output if the block has one.
+
+    Handles singular (qwen2_moe: ``shared_expert``), plural
+    (glm_moe_dsa/deepseek_v3: ``shared_experts``), and MLP
+    (hunyuan_v1_moe: ``shared_mlp``) attribute names.
+
+    peft wraps individual linear layers inside the shared expert with
+    standard LoRA — calling forward() handles this transparently.
+    """
+    shared_expert = (
+        getattr(moe_block, "shared_expert", None)
+        or getattr(moe_block, "shared_experts", None)
+        or getattr(moe_block, "shared_mlp", None)
+    )
+    if shared_expert is None:
+        return None
+
+    shared_expert_output = shared_expert(hidden_states_flat)
+
+    # Optional sigmoid gate (Qwen2MoE pattern).
+    # shared_expert_gate may also be peft-wrapped (standard LoRA
+    # on nn.Linear), its forward() applies LoRA automatically.
+    shared_expert_gate = getattr(moe_block, "shared_expert_gate", None)
+    if shared_expert_gate is not None:
+        shared_expert_output = (
+            F.sigmoid(shared_expert_gate(hidden_states_flat)) * shared_expert_output
+        )
+
+    return shared_expert_output
+
+
 # =============================================================================
 # Layer classes
 # =============================================================================
@@ -281,16 +433,18 @@ class ScatterMoEGatedMLP(nn.Module):
 
 class HFScatterMoEGatedMLP(nn.Module):
     """
-    ScatterMoE-accelerated forward pass for HF MoEs (OLMoE / Qwen2MoE).
+    ScatterMoE-accelerated forward pass for HF MoEs.
 
     Used as a kernel layer via the HF ``kernels`` library.  The ``forward``
-    method replaces the original ``OlmoeSparseMoeBlock.forward``.
+    method replaces the original SparseMoeBlock.forward.
 
-    Supports both full-parameter training and LoRA fine-tuning:
+    Supports:
 
-    * **Full-param**: uses ``parallel_linear`` (base ScatterMoE kernel)
-    * **LoRA**: detects peft ``ParamWrapper`` on ``self.experts``, extracts
-      adapter weights, and uses ``parallel_linear_lora`` (fused kernel)
+    * **Softmax→topk routing**: OLMoE, Qwen2/3MoE, Mixtral, MiniMax
+    * **Sigmoid→topk routing**: GLM, DeepSeek V3, MiniMax M2
+    * **Full-parameter training**: uses ``parallel_linear`` (base ScatterMoE)
+    * **LoRA fine-tuning**: detects peft ``ParamWrapper`` on ``self.experts``,
+      extracts adapter weights, and uses ``parallel_linear_lora`` (fused kernel)
     """
 
     @staticmethod
@@ -302,7 +456,7 @@ class HFScatterMoEGatedMLP(nn.Module):
             self: The MoeSparseMoeBlock module containing:
                 - self.gate: Router (or peft ParamWrapper wrapping it)
                 - self.experts: Experts module (or peft ParamWrapper chain)
-                - self.shared_expert: Optional shared expert (e.g. Qwen2MoE)
+                - self.shared_expert(s): Optional shared expert
                 - self.shared_expert_gate: Optional shared expert gate
             layer_input: Input tensor [batch_size, seq_len, hidden_size]
 
@@ -313,38 +467,17 @@ class HFScatterMoEGatedMLP(nn.Module):
         hidden_states_flat = layer_input.view(-1, hidden_dim)
 
         # ====================================================================
-        # Shared Expert (if present, e.g. Qwen2MoE)
+        # Shared Expert (if present, e.g. Qwen2MoE, DeepSeek V3)
         # ====================================================================
-        # peft wraps individual linear layers inside shared_expert with
-        # standard LoRA — calling forward() handles this transparently.
-        if hasattr(self, "shared_expert") and self.shared_expert is not None:
-            shared_expert_output = self.shared_expert(hidden_states_flat)
-            # shared_expert_gate may also be peft-wrapped (standard LoRA
-            # on nn.Linear), its forward() applies LoRA automatically.
-            shared_expert_gate_output = F.sigmoid(
-                self.shared_expert_gate(hidden_states_flat)
-            )
-            shared_expert_output = shared_expert_output * shared_expert_gate_output
-        else:
-            shared_expert_output = None
+        shared_expert_output = _compute_shared_expert(self, hidden_states_flat)
 
         # ====================================================================
         # Router Computation (with optional gate LoRA)
         # ====================================================================
         base_gate, gate_weight, gate_lora_delta = _unwrap_gate_lora(self.gate)
-        router_logits = F.linear(hidden_states_flat, gate_weight)
-        if gate_lora_delta is not None:
-            router_logits = router_logits + F.linear(
-                hidden_states_flat, gate_lora_delta
-            )
-        routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
-
-        top_k = base_gate.top_k
-        num_experts = base_gate.num_experts
-        routing_weights, selected_experts = torch.topk(routing_weights, top_k, dim=-1)
-
-        if base_gate.norm_topk_prob:
-            routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
+        routing_weights, selected_experts, top_k, num_experts = _route(
+            self, base_gate, hidden_states_flat, gate_weight, gate_lora_delta
+        )
         routing_weights = routing_weights.to(hidden_states_flat.dtype)
 
         sorted_expert_idxs, sorted_scattered_idxs, expert_offsets = flatten_sort_count(
@@ -356,20 +489,71 @@ class HFScatterMoEGatedMLP(nn.Module):
         # ====================================================================
         experts, gup_lora, down_lora = _unwrap_experts_lora(self.experts)
 
+        # ====================================================================
+        # Selective expert weight dequantization
+        # ====================================================================
+        # When experts are BnB-quantized (quantize_moe_experts), dequantize
+        # only the active experts instead of all E. This saves ~97% memory
+        # for the transient dequant buffer when few experts are active.
+        use_selective = (
+            getattr(self, "_use_selective_dequant", False)
+            and hasattr(experts, "parametrizations")
+            and "gate_up_proj" in experts.parametrizations
+        )
+
+        if use_selective:
+            from axolotl.integrations.kernels.libs.scattermoe_lora.selective_dequant import (
+                get_active_experts,
+                remap_expert_indices,
+                selective_expert_weights,
+                selective_lora_weights,
+            )
+
+            active_experts = get_active_experts(sorted_expert_idxs, num_experts)
+            remapped_expert_idxs, compact_offsets = remap_expert_indices(
+                sorted_expert_idxs,
+                expert_offsets,
+                active_experts,
+                num_experts,
+            )
+            # Dequantize only active experts' weights
+            gate_up_W = selective_expert_weights(
+                experts,
+                "gate_up_proj",
+                active_experts,
+            ).transpose(2, 1)  # [num_active, hidden, 2*inter]
+
+            # Remap LoRA weights to match compact expert indices
+            if gup_lora is not None:
+                gup_A, gup_B, gup_scaling = gup_lora
+                gup_A, gup_B = selective_lora_weights(
+                    gup_A,
+                    gup_B,
+                    active_experts,
+                    num_experts,
+                )
+                gup_lora = (gup_A, gup_B, gup_scaling)
+
+            # Use remapped indices for ScatterMoE kernels
+            sei_gup = remapped_expert_idxs
+            eo_gup = compact_offsets
+        else:
+            gate_up_W = experts.gate_up_proj.transpose(2, 1)  # [E, hidden, 2*inter]
+            sei_gup = sorted_expert_idxs
+            eo_gup = expert_offsets
+
         # ====================================================================
         # Gate + Up projection
         # ====================================================================
-        gate_up_W = experts.gate_up_proj.transpose(2, 1)  # [E, hidden, 2*inter]
-
         if gup_lora is not None:
             gup_A, gup_B, gup_scaling = gup_lora
             gup = parallel_linear_lora(
                 hidden_states_flat,
                 gate_up_W,
                 top_k,
-                sorted_expert_idxs,
+                sei_gup,
                 sorted_scattered_idxs,
-                expert_offsets,
+                eo_gup,
                 lora_A=gup_A,
                 lora_B=gup_B,
                 scaling=gup_scaling,
@@ -383,9 +567,9 @@ class HFScatterMoEGatedMLP(nn.Module):
                 hidden_states_flat,
                 gate_up_W,
                 top_k,
-                sorted_expert_idxs,
+                sei_gup,
                 sorted_scattered_idxs,
-                expert_offsets,
+                eo_gup,
                 grouped_in=False,
                 grouped_out=True,
             )
@@ -396,7 +580,29 @@ class HFScatterMoEGatedMLP(nn.Module):
         # ====================================================================
         # Down projection
         # ====================================================================
-        down_W = experts.down_proj.transpose(2, 1)  # [E, inter, hidden]
+        if use_selective:
+            down_W = selective_expert_weights(
+                experts,
+                "down_proj",
+                active_experts,
+            ).transpose(2, 1)  # [num_active, inter, hidden]
+
+            if down_lora is not None:
+                down_A, down_B, down_scaling = down_lora
+                down_A, down_B = selective_lora_weights(
+                    down_A,
+                    down_B,
+                    active_experts,
+                    num_experts,
+                )
+                down_lora = (down_A, down_B, down_scaling)
+
+            sei_down = remapped_expert_idxs
+            eo_down = compact_offsets
+        else:
+            down_W = experts.down_proj.transpose(2, 1)  # [E, inter, hidden]
+            sei_down = sorted_expert_idxs
+            eo_down = expert_offsets
 
         if down_lora is not None:
             down_A, down_B, down_scaling = down_lora
@@ -404,9 +610,9 @@ class HFScatterMoEGatedMLP(nn.Module):
                 h,
                 down_W,
                 1,
-                sorted_expert_idxs,
+                sei_down,
                 sorted_scattered_idxs,
-                expert_offsets,
+                eo_down,
                 lora_A=down_A,
                 lora_B=down_B,
                 scaling=down_scaling,
@@ -421,9 +627,9 @@ class HFScatterMoEGatedMLP(nn.Module):
                 h,
                 down_W,
                 1,
-                sorted_expert_idxs,
+                sei_down,
                 sorted_scattered_idxs,
-                expert_offsets,
+                eo_down,
                 grouped_in=True,
                 grouped_out=False,
                 gates=routing_weights,

src/axolotl/integrations/kernels/libs/scattermoe_lora/selective_dequant.py

@@ -0,0 +1,282 @@
+"""
+Selective Expert Dequantization
+===============================
+
+Instead of dequantizing all E expert weight matrices at once (which creates
+a ~1 GB transient buffer for 256 experts), only dequantize the experts that
+are actually routed to by the current batch's top-k selection.
+
+For Qwen3.5-35B-A3B (E=256, top_k=8, hidden=2048, intermediate=512):
+  - Full dequant: [256, 2048, 1024] = 1,074 MB per projection
+  - Selective (8 active): [8, 2048, 1024] = 33.5 MB per projection
+  - Savings: ~97% memory reduction per layer
+
+This module provides format-agnostic selective weight extraction:
+  - BnB 4-bit (nf4/fp4): slice quantized data + absmax per expert
+  - bf16/fp32: direct indexing (no dequant needed)
+  - FP8: slice + cast
+
+The ScatterMoE kernel itself doesn't change — we remap expert indices
+from global (0..E-1) to compact (0..num_active-1) and pass the smaller
+weight tensor.
+"""
+
+import torch
+import torch.nn as nn
+
+
+def get_active_experts(sorted_expert_idxs: torch.Tensor, E: int) -> torch.Tensor:
+    """Get sorted unique expert indices from the routing output.
+
+    Args:
+        sorted_expert_idxs: Expert assignments sorted by expert id [T*k]
+        E: Total number of experts
+
+    Returns:
+        active: Sorted unique expert indices [num_active]
+    """
+    return torch.unique(sorted_expert_idxs)
+
+
+def remap_expert_indices(
+    sorted_expert_idxs: torch.Tensor,
+    expert_offsets: torch.Tensor,
+    active_experts: torch.Tensor,
+    E: int,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    """Remap global expert indices to compact indices.
+
+    Maps expert ids from [0..E-1] to [0..num_active-1], preserving the
+    sort order. Also compacts expert_offsets to only active experts.
+
+    Args:
+        sorted_expert_idxs: [T*k] expert ids in sorted order
+        expert_offsets: [E] cumulative token counts (original)
+        active_experts: [num_active] sorted unique expert ids
+        E: Total number of experts
+
+    Returns:
+        remapped_idxs: [T*k] expert ids in [0..num_active-1]
+        compact_offsets: [num_active] cumulative token counts
+    """
+    # Build remap table: global_id -> compact_id
+    remap = torch.empty(E, dtype=torch.long, device=sorted_expert_idxs.device)
+    remap[active_experts] = torch.arange(
+        len(active_experts), device=sorted_expert_idxs.device
+    )
+
+    remapped_idxs = remap[sorted_expert_idxs]
+
+    # Compact the expert_offsets: only keep active experts' cumulative counts
+    compact_offsets = expert_offsets[active_experts]
+
+    return remapped_idxs, compact_offsets
+
+
+def _selective_dequant_bnb4(
+    raw_param: torch.Tensor,
+    quant_state,
+    active_experts: torch.Tensor,
+    expert_shape: tuple[int, int],
+) -> torch.Tensor:
+    """Dequantize only selected experts from BnB 4-bit packed data.
+
+    The raw parameter is a flattened 4-bit packed tensor. Each expert's
+    data is contiguous (stored in expert-major order), so we can gather
+    the packed data and absmax blocks for active experts, then dequantize
+    as one contiguous block.
+
+    Args:
+        raw_param: Flattened uint8 tensor of packed 4-bit weights
+        quant_state: BnB QuantState with absmax, blocksize, code, etc.
+        active_experts: [num_active] expert indices to dequantize
+        expert_shape: (dim1, dim2) shape per expert (e.g. (1024, 2048))
+
+    Returns:
+        Dequantized weights [num_active, dim1, dim2] in original dtype
+    """
+    import bitsandbytes.functional as F  # noqa: N812
+    from bitsandbytes.functional import QuantState
+
+    expert_numel = expert_shape[0] * expert_shape[1]
+    packed_per_expert = expert_numel // 2  # 4-bit = 2 values per byte
+    blocks_per_expert = expert_numel // quant_state.blocksize
+    num_active = len(active_experts)
+
+    if blocks_per_expert == 0:
+        # Expert is smaller than one quantization block — blocks span across
+        # expert boundaries, so per-expert slicing isn't possible.
+        # Fallback: full dequantize + index.
+        full = F.dequantize_4bit(raw_param, quant_state)
+        E_total = full.numel() // expert_numel
+        return full.reshape(E_total, *expert_shape)[active_experts]
+
+    # Use fused Triton kernel for NF4 (handles selective gather + dequant in one pass)
+    if quant_state.quant_type == "nf4" and raw_param.dtype == torch.uint8:
+        from axolotl.integrations.kernels.libs.scattermoe_lora.selective_dequant_kernel import (
+            selective_dequant_nf4_triton,
+        )
+
+        # Handle nested (double) quantization: dequantize absmax first
+        # BnB uses dequantize_blockwise (not _4bit) for nested absmax + offset
+        if quant_state.nested:
+            absmax = F.dequantize_blockwise(quant_state.absmax, quant_state.state2)
+            absmax += quant_state.offset
+            if absmax.dtype != torch.float32:
+                absmax = absmax.float()
+        else:
+            absmax = quant_state.absmax
+
+        return selective_dequant_nf4_triton(
+            packed_data=raw_param,
+            absmax=absmax,
+            active_experts=active_experts,
+            expert_shape=expert_shape,
+            blocksize=quant_state.blocksize,
+            dtype=quant_state.dtype,
+            codebook=quant_state.code,
+        )
+
+    # Fallback: gather + BnB dequant (for fp4 or non-uint8 packed formats)
+    raw_flat = raw_param.reshape(-1)
+
+    offsets_qt = (
+        active_experts.long()[:, None] * packed_per_expert
+        + torch.arange(packed_per_expert, device=raw_param.device)[None, :]
+    ).reshape(-1)
+    qt_gathered = raw_flat[offsets_qt]
+
+    offsets_abs = (
+        active_experts.long()[:, None] * blocks_per_expert
+        + torch.arange(blocks_per_expert, device=raw_param.device)[None, :]
+    ).reshape(-1)
+
+    if quant_state.nested:
+        full_absmax = F.dequantize_blockwise(quant_state.absmax, quant_state.state2)
+        full_absmax += quant_state.offset
+        if full_absmax.dtype != torch.float32:
+            full_absmax = full_absmax.float()
+        absmax_gathered = full_absmax[offsets_abs]
+    else:
+        absmax_gathered = quant_state.absmax[offsets_abs]
+
+    qt_gathered = qt_gathered.unsqueeze(1) if qt_gathered.dim() == 1 else qt_gathered
+
+    gathered_qs = QuantState(
+        absmax=absmax_gathered,
+        shape=torch.Size([num_active * expert_numel]),
+        blocksize=quant_state.blocksize,
+        quant_type=quant_state.quant_type,
+        code=quant_state.code,
+        dtype=quant_state.dtype,
+    )
+
+    deq = F.dequantize_4bit(qt_gathered, gathered_qs)
+    return deq.reshape(num_active, *expert_shape)
+
+
+def _selective_index_dense(
+    param: torch.Tensor,
+    active_experts: torch.Tensor,
+) -> torch.Tensor:
+    """Select experts from a dense (bf16/fp32) weight tensor.
+
+    Simple indexing — no dequantization needed.
+    """
+    return param[active_experts]
+
+
+def selective_expert_weights(
+    experts_module: nn.Module,
+    param_name: str,
+    active_experts: torch.Tensor,
+) -> torch.Tensor:
+    """Extract and dequantize only the active experts' weights.
+
+    Format-agnostic: dispatches based on whether the parameter is
+    BnB 4-bit quantized (via parametrize), FP8, or dense bf16/fp32.
+
+    Args:
+        experts_module: The base experts module (e.g. Qwen3_5MoeExperts)
+        param_name: "gate_up_proj" or "down_proj"
+        active_experts: [num_active] sorted unique expert indices
+
+    Returns:
+        Compact weight tensor [num_active, dim1, dim2] ready for ScatterMoE
+    """
+    # Check if the parameter is BnB-quantized via parametrize
+    if (
+        hasattr(experts_module, "parametrizations")
+        and param_name in experts_module.parametrizations
+    ):
+        param_list = experts_module.parametrizations[param_name]
+        parametrization = param_list[0]
+
+        # BnB 4-bit parametrization
+        if hasattr(parametrization, "quant_state"):
+            # The raw quantized data is on the ParametrizationList, not the
+            # individual Bnb4bitParametrization module
+            raw_param = param_list.original
+            qs = parametrization.quant_state
+            # qs.shape is the original tensor shape before flattening.
+            # For MoE experts it's [E, d1, d2] (3D) or [total_elements] (1D).
+            orig_shape = qs.shape
+            if isinstance(orig_shape, torch.Size) and len(orig_shape) == 3:
+                expert_shape = (orig_shape[1], orig_shape[2])
+            elif isinstance(orig_shape, torch.Size) and len(orig_shape) == 1:
+                # Flattened — need to infer from module attributes
+                E_total = getattr(experts_module, "num_experts", None)
+                if E_total is None:
+                    E_total = int(active_experts.max().item()) + 1
+                expert_numel = orig_shape[0] // E_total
+                d2 = getattr(experts_module, "hidden_dim", None) or getattr(
+                    experts_module, "intermediate_dim", None
+                )
+                if d2 and expert_numel % d2 == 0:
+                    expert_shape = (expert_numel // d2, d2)
+                else:
+                    full = getattr(experts_module, param_name)
+                    return full[active_experts]
+            else:
+                full = getattr(experts_module, param_name)
+                return full[active_experts]
+
+            return _selective_dequant_bnb4(raw_param, qs, active_experts, expert_shape)
+
+    # Dense parameter (bf16/fp32) — direct indexing
+    param = getattr(experts_module, param_name)
+    if param.dim() == 3:
+        return param[active_experts]
+
+    # Fallback: full access
+    return param
+
+
+def selective_lora_weights(
+    lora_A: torch.Tensor,
+    lora_B: torch.Tensor,
+    active_experts: torch.Tensor,
+    E: int,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    """Select LoRA A and B weights for only the active experts.
+
+    LoRA layout (scattermoe format):
+      A: [r*E, K] — expert e occupies rows [e*r : (e+1)*r]
+      B: [N, r*E] — expert e occupies cols [e*r : (e+1)*r]
+
+    Returns compact:
+      A: [r*num_active, K]
+      B: [N, r*num_active]
+    """
+    R = lora_A.size(0) // E
+
+    # Vectorized gather: active_experts[:, None] * R + arange(R)[None, :]
+    row_idx = (
+        active_experts.long()[:, None] * R
+        + torch.arange(R, device=lora_A.device)[None, :]
+    ).reshape(-1)
+
+    compact_A = lora_A[row_idx]  # [r*num_active, K]
+    compact_B = lora_B[:, row_idx]  # [N, r*num_active]
+
+    return compact_A, compact_B

src/axolotl/integrations/kernels/libs/scattermoe_lora/selective_dequant_kernel.py

@@ -0,0 +1,179 @@
+"""
+Triton kernel for fused selective expert gather + NF4 dequantization.
+
+Instead of:
+  1. Gather packed uint8 data for active experts (memory copy)
+  2. Gather absmax for active experts (memory copy)
+  3. Call BnB dequantize_4bit CUDA kernel
+
+This kernel does all three in one pass:
+  - Reads packed NF4 bytes from expert-strided positions
+  - Looks up the NF4 codebook
+  - Multiplies by the per-block absmax
+  - Writes bf16 output directly
+
+This eliminates the intermediate gather buffer entirely.
+"""
+
+import torch
+import triton
+import triton.language as tl
+
+# NF4 codebook (16 values, precomputed by BnB)
+# These are the normalized float4 reconstruction values
+NF4_CODEBOOK = [
+    -1.0,
+    -0.6961928009986877,
+    -0.5250730514526367,
+    -0.39491748809814453,
+    -0.28444138169288635,
+    -0.18477343022823334,
+    -0.09105003625154495,
+    0.0,
+    0.07958029955625534,
+    0.16093020141124725,
+    0.24611230194568634,
+    0.33791524171829224,
+    0.44070982933044434,
+    0.5626170039176941,
+    0.7229568362236023,
+    1.0,
+]
+
+
+@triton.jit
+def _selective_dequant_nf4_kernel(
+    # Input: packed NF4 data (flattened, expert-major order)
+    packed_ptr,
+    # Input: absmax values (flattened, expert-major order)
+    absmax_ptr,
+    # Input: active expert indices
+    active_experts_ptr,
+    # Input: NF4 codebook (16 float values)
+    codebook_ptr,
+    # Output: dequantized bf16 weights [num_active, expert_numel]
+    out_ptr,
+    stride_out_e,  # stride for expert dim in output
+    # Dimensions
+    num_active,
+    packed_per_expert,  # expert_numel // 2
+    blocks_per_expert,  # expert_numel // blocksize
+    blocksize: tl.constexpr,
+    # Tile size
+    BLOCK_SIZE: tl.constexpr,  # elements per thread block (must be multiple of 2)
+):
+    """
+    Each program processes BLOCK_SIZE elements from one expert.
+
+    Grid: (num_active, cdiv(expert_numel, BLOCK_SIZE))
+
+    For each output element:
+      1. Compute which byte in packed data contains this element
+      2. Extract the 4-bit nibble (high or low)
+      3. Look up in NF4 codebook
+      4. Scale by absmax for this block
+    """
+    expert_local_idx = tl.program_id(0)  # which active expert (0..num_active-1)
+    block_id = tl.program_id(1)  # which element block
+
+    # Load the global expert index
+    expert_global = tl.load(active_experts_ptr + expert_local_idx).to(tl.int64)
+
+    expert_numel = packed_per_expert * 2  # 2 elements per packed byte
+    elem_offset = block_id * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
+    mask = elem_offset < expert_numel
+
+    # Each element is packed as: byte[i//2], low nibble for even i, high for odd i
+    byte_idx = elem_offset // 2
+    is_high = (elem_offset % 2) == 1
+
+    # Read packed bytes from the global expert's region
+    packed_global_offset = expert_global * packed_per_expert + byte_idx
+    packed_bytes = tl.load(packed_ptr + packed_global_offset, mask=mask, other=0).to(
+        tl.int32
+    )
+
+    # Extract 4-bit nibble
+    # BnB packing: high nibble = even element, low nibble = odd element
+    nibble = tl.where(is_high, packed_bytes & 0xF, (packed_bytes >> 4) & 0xF)
+
+    # NF4 codebook lookup
+    # Load all 16 codebook values (small, fits in registers)
+    # Use gather from codebook pointer
+    code_val = tl.load(codebook_ptr + nibble, mask=mask, other=0.0)
+
+    # Load absmax for this element's quantization block
+    block_idx = elem_offset // blocksize
+    absmax_global_offset = expert_global * blocks_per_expert + block_idx
+    absmax_val = tl.load(absmax_ptr + absmax_global_offset, mask=mask, other=1.0)
+
+    # Dequantize: value = codebook[nibble] * absmax
+    result = code_val * absmax_val
+
+    # Store to output
+    out_offset = expert_local_idx * stride_out_e + elem_offset
+    tl.store(out_ptr + out_offset, result.to(out_ptr.dtype.element_ty), mask=mask)
+
+
+def selective_dequant_nf4_triton(
+    packed_data: torch.Tensor,
+    absmax: torch.Tensor,
+    active_experts: torch.Tensor,
+    expert_shape: tuple[int, int],
+    blocksize: int,
+    dtype: torch.dtype = torch.bfloat16,
+    codebook: torch.Tensor | None = None,
+) -> torch.Tensor:
+    """Fused selective gather + NF4 dequantization via Triton kernel.
+
+    Args:
+        packed_data: Flattened packed NF4 data [total_packed] or [total_packed, 1]
+        absmax: Per-block scaling factors [total_blocks]
+        active_experts: Sorted indices of experts to dequantize [num_active]
+        expert_shape: (dim1, dim2) per expert
+        blocksize: Quantization block size
+        dtype: Output dtype (default bf16)
+        codebook: NF4 lookup table [16] (uses default NF4 codebook if None)
+
+    Returns:
+        Dequantized weights [num_active, dim1, dim2]
+    """
+    num_active = active_experts.shape[0]
+    expert_numel = expert_shape[0] * expert_shape[1]
+    packed_per_expert = expert_numel // 2
+    blocks_per_expert = expert_numel // blocksize
+
+    # Prepare codebook on device
+    if codebook is None:
+        codebook = torch.tensor(
+            NF4_CODEBOOK, dtype=torch.float32, device=packed_data.device
+        )
+    else:
+        codebook = codebook.to(device=packed_data.device, dtype=torch.float32)
+
+    # Flatten inputs
+    packed_flat = packed_data.reshape(-1)
+    absmax_flat = absmax.reshape(-1).float()  # absmax is usually fp32
+
+    # Output buffer
+    out = torch.empty(num_active, expert_numel, dtype=dtype, device=packed_data.device)
+
+    BLOCK_SIZE = 1024  # Process 1024 elements per thread block
+
+    grid = (num_active, triton.cdiv(expert_numel, BLOCK_SIZE))
+
+    _selective_dequant_nf4_kernel[grid](
+        packed_flat,
+        absmax_flat,
+        active_experts,
+        codebook,
+        out,
+        out.stride(0),
+        num_active=num_active,
+        packed_per_expert=packed_per_expert,
+        blocks_per_expert=blocks_per_expert,
+        blocksize=blocksize,
+        BLOCK_SIZE=BLOCK_SIZE,
+    )
+
+    return out.reshape(num_active, *expert_shape)

src/axolotl/integrations/kernels/plugin.py

@@ -61,9 +61,20 @@ class KernelsPlugin(BasePlugin):
         return "axolotl.integrations.kernels.KernelsArgs"
 
     def pre_model_load(self, cfg):
+        from axolotl.integrations.kernels.constants import SPARSE_MOE_BLOCK
+
+        # Prefer text backbone type for VLMs, but fall back to base type
+        # when the text type isn't in the supported mapping (e.g. qwen3_5_moe_text)
+        moe_model_type = cfg.model_config_type_text or cfg.model_config_type
+        if (
+            moe_model_type not in SPARSE_MOE_BLOCK
+            and cfg.model_config_type in SPARSE_MOE_BLOCK
+        ):
+            moe_model_type = cfg.model_config_type
+
         if cfg.use_scattermoe:
             self._register_kernels()
-            self._kernelize_model(cfg.model_config_type)
+            self._kernelize_model(moe_model_type)
         elif cfg.use_sonicmoe:
             if not importlib.util.find_spec("sonicmoe"):
                 raise RuntimeError(
@@ -75,11 +86,9 @@ class KernelsPlugin(BasePlugin):
 
             from axolotl.integrations.kernels.sonicmoe import patch_sonicmoe
 
-            LOG.info(
-                f"Applying SonicMoE patches for model type: {cfg.model_config_type}"
-            )
+            LOG.info(f"Applying SonicMoE patches for model type: {moe_model_type}")
             patch_sonicmoe(
-                cfg.model_config_type,
+                moe_model_type,
                 torch_compile=bool(getattr(cfg, "torch_compile", False)),
             )
 
@@ -110,6 +119,16 @@ class KernelsPlugin(BasePlugin):
             }
         )
 
+    def add_callbacks_pre_trainer(self, cfg, model):
+        callbacks = []
+        if cfg.use_scattermoe:
+            from axolotl.integrations.kernels.autotune_callback import (
+                AutotuneReportCallback,
+            )
+
+            callbacks.append(AutotuneReportCallback())
+        return callbacks
+
     def _kernelize_model(self, model_type: str):
         from kernels import replace_kernel_forward_from_hub
 

src/axolotl/integrations/kernels/sonicmoe/routing.py

@@ -5,6 +5,7 @@ Different MoE architectures use different routing strategies:
 - qwen3_moe / qwen2_moe / qwen3_5_moe / qwen3_vl_moe / qwen3_omni_moe: softmax -> topk (with optional renormalization)
 - gpt_oss: topk -> softmax (uses fused moe_TC_softmax_topk_layer, routing_fn=None)
 - glm_moe_dsa: sigmoid -> topk (with group-based expert selection)
+- mistral4: softmax -> group selection -> topk (with renormalization and scaling)
 
 Each model type maps to a (routing_fn, activation_type, router_attr) triple.
 When routing_fn is None, the fused moe_TC_softmax_topk_layer path is used.
@@ -45,6 +46,8 @@ def get_model_moe_config(model_type: str):
         "minimax",
     ):
         return softmax_topk_routing, ActivationType.SWIGLU, "gate"
+    elif model_type in ("mistral4",):
+        return softmax_group_topk_routing, ActivationType.SWIGLU, "gate"
     elif model_type in (
         "glm_moe_dsa",
         "deepseek_v3",
@@ -126,6 +129,62 @@ def softmax_topk_routing(
     return flat_scores, flat_token_idx, flat_expert_idx, router_logits
 
 
+def softmax_group_topk_routing(
+    hidden_states: torch.Tensor, moe_block
+) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+    """Mistral4-style routing: softmax -> group selection -> topk -> renorm -> scale."""
+    gate = moe_block.gate
+    T, H = hidden_states.shape
+    K = moe_block.top_k
+    E = getattr(moe_block, "n_routed_experts", gate.weight.shape[0])
+    n_group = getattr(moe_block, "n_group", 1)
+
+    router_logits = F.linear(hidden_states, gate.weight)  # [T, E]
+    router_probs = F.softmax(router_logits, dim=-1, dtype=torch.float32)  # [T, E]
+
+    scores_for_choice = router_probs
+
+    # Group selection: pick top groups, mask the rest
+    if n_group > 1:
+        group_scores = (
+            scores_for_choice.view(-1, n_group, E // n_group)
+            .topk(2, dim=-1)[0]
+            .sum(dim=-1)
+        )
+        group_idx = torch.topk(
+            group_scores, k=moe_block.topk_group, dim=-1, sorted=False
+        )[1]
+        group_mask = torch.zeros_like(group_scores)
+        group_mask.scatter_(1, group_idx, 1)
+        score_mask = (
+            group_mask.unsqueeze(-1).expand(-1, n_group, E // n_group).reshape(-1, E)
+        )
+        scores_for_choice = scores_for_choice.masked_fill(~score_mask.bool(), 0.0)
+
+    topk_indices = torch.topk(scores_for_choice, k=K, dim=-1, sorted=False)[1]
+    topk_weights = router_probs.gather(1, topk_indices)
+
+    # Renormalization + scaling
+    norm_topk_prob = getattr(moe_block, "norm_topk_prob", True)
+    if norm_topk_prob:
+        topk_weights = topk_weights / (topk_weights.sum(dim=-1, keepdim=True) + 1e-20)
+    routed_scaling_factor = getattr(moe_block, "routed_scaling_factor", 1.0)
+    topk_weights = topk_weights * routed_scaling_factor
+
+    # Flatten for moe_general_routing_inputs
+    token_indices = (
+        torch.arange(T, device=hidden_states.device, dtype=torch.int32)
+        .unsqueeze(1)
+        .expand(T, K)
+    )
+
+    flat_scores = topk_weights.to(torch.float32).reshape(-1)  # [T*K]
+    flat_token_idx = token_indices.reshape(-1)  # [T*K]
+    flat_expert_idx = topk_indices.to(torch.int32).reshape(-1)  # [T*K]
+
+    return flat_scores, flat_token_idx, flat_expert_idx, router_logits
+
+
 def sigmoid_topk_routing(
     hidden_states: torch.Tensor, moe_block
 ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:

src/axolotl/kernels/lora.py

@@ -25,7 +25,7 @@ def get_lora_parameters(
 ) -> tuple[
     torch.Tensor,
     torch.Tensor | None,
-    QuantState | None,
+    QuantState | torch.Tensor | None,
     torch.Tensor | None,
     torch.Tensor | None,
     float | None,
@@ -48,9 +48,13 @@ def get_lora_parameters(
 
     if not hasattr(proj, "disable_adapters") or proj.disable_adapters or proj.merged:
         quant_state = getattr(W, "quant_state", None)
+        if quant_state is None and W.dtype == torch.float8_e4m3fn:
+            quant_state = getattr(base_layer, "weight_scale_inv", None)
         return W, b, quant_state, None, None, None
 
     quant_state = getattr(W, "quant_state", None)
+    if quant_state is None and W.dtype == torch.float8_e4m3fn:
+        quant_state = getattr(base_layer, "weight_scale_inv", None)
 
     active_adapter = (
         proj.active_adapters[0]
@@ -81,7 +85,7 @@ def matmul_lora(
     X: torch.Tensor,
     W: torch.Tensor,
     b: torch.Tensor | None,
-    W_quant: QuantState | None,
+    W_quant: QuantState | torch.Tensor | None,
     A: torch.Tensor | None,
     B: torch.Tensor | None,
     s: float | None,
@@ -636,7 +640,9 @@ class LoRA_QKV(torch.autograd.Function):
         del q_weight
         del q_weight_t
         if A_q is not None and B_q is not None:
-            grad_X.addmm_(q_grad, torch.mm(B_q_scaled, A_q_scaled))
+            # Stay decomposed: dQ @ B^T gives [T, R], then [T, R] @ (s*A) gives [T, in]
+            # This is 65x fewer FLOPs than materializing B@A into [out, in]
+            grad_X.addmm_(torch.mm(q_grad, B_q_scaled), A_q_scaled)
 
         # K path
         k_weight_t = dequantize(k_weight, k_quant)
@@ -644,7 +650,7 @@ class LoRA_QKV(torch.autograd.Function):
         del k_weight
         del k_weight_t
         if A_k is not None and B_k is not None:
-            grad_X.addmm_(k_grad, torch.mm(B_k_scaled, A_k_scaled))
+            grad_X.addmm_(torch.mm(k_grad, B_k_scaled), A_k_scaled)
 
         # V path
         v_weight_t = dequantize(v_weight, v_quant)
@@ -652,7 +658,7 @@ class LoRA_QKV(torch.autograd.Function):
         del v_weight
         del v_weight_t
         if A_v is not None and B_v is not None:
-            grad_X.addmm_(v_grad, torch.mm(B_v_scaled, A_v_scaled))
+            grad_X.addmm_(torch.mm(v_grad, B_v_scaled), A_v_scaled)
 
         # Transpose gradients if needed
         if d_A_q is not None:
@@ -815,7 +821,8 @@ class LoRA_O(torch.autograd.Function):
         del W
 
         A, B = A.to(dtype), B.to(dtype)
-        dX += s * dY @ B @ A
+        # Stay decomposed: dY @ B gives [T, R], then [T, R] @ A gives [T, in]
+        dX.addmm_(torch.mm(dY, B), A, alpha=s)
 
         # W, b, W_quant, A, B, s
         return dX.view(batch, seq_len, hd), None, None, None, d_A.t(), d_B.t(), None

src/axolotl/kernels/quantize.py

@@ -1,4 +1,4 @@
-"""Dequantization utilities for `bitsandbytes` integration."""
+"""Dequantization utilities for `bitsandbytes` and FP8 integration."""
 
 import ctypes
 
@@ -15,9 +15,50 @@ CUDA_STREAM: torch.cuda.Stream | None = None
 HAS_CUDA_STREAM: bool = Version(bnb.__version__) > Version("0.43.3")
 
 
+def dequantize_fp8(
+    W: torch.Tensor,
+    scale_inv: torch.Tensor,
+    dtype: torch.dtype = torch.bfloat16,
+) -> torch.Tensor:
+    """Dequantize FP8 block-quantized weights: W_dequant = W_fp8 * scale_inv.
+
+    Args:
+        W: FP8 weight tensor [out_features, in_features] in float8_e4m3fn.
+        scale_inv: Per-block inverse scale [ceil(out/block), ceil(in/block)]
+            or per-tensor scalar.
+        dtype: Output dtype (default bf16).
+
+    Returns:
+        Dequantized tensor in the specified dtype.
+    """
+    W_float = W.to(dtype)
+    if scale_inv.numel() == 1:
+        return W_float * scale_inv.to(dtype)
+    if scale_inv.dim() == 2 and W.dim() == 2:
+        sr, sc = scale_inv.shape
+        br = W.shape[0] // sr
+        bc = W.shape[1] // sc
+        # If dimensions are exactly divisible, use fast reshape path
+        if sr * br == W.shape[0] and sc * bc == W.shape[1]:
+            return (
+                W_float.reshape(sr, br, sc, bc) * scale_inv[:, None, :, None].to(dtype)
+            ).reshape(W.shape)
+        # Tail-block handling: compute actual block size (ceil division),
+        # tile scale_inv to cover full shape, then crop to W's dimensions
+        br_ceil = -(-W.shape[0] // sr)  # ceil(rows / scale_rows) = block_size
+        bc_ceil = -(-W.shape[1] // sc)
+        scale_expanded = (
+            scale_inv.to(dtype)
+            .repeat_interleave(br_ceil, dim=0)
+            .repeat_interleave(bc_ceil, dim=1)
+        )[: W.shape[0], : W.shape[1]]
+        return W_float * scale_expanded
+    return W_float * scale_inv.to(dtype)
+
+
 def dequantize(
     W: torch.Tensor,
-    quant_state: QuantState | list | None = None,
+    quant_state: QuantState | list | torch.Tensor | None = None,
     out: torch.Tensor | None = None,
 ) -> torch.Tensor:
     """
@@ -49,6 +90,15 @@ def dequantize(
     if quant_state is None:
         return W
 
+    # FP8 path: quant_state is actually scale_inv tensor
+    if W.dtype == torch.float8_e4m3fn:
+        scale_inv = quant_state
+        # Caller may pass W.t() (non-contiguous) — dequantize in original
+        # layout then transpose back so the result shape matches the input.
+        if not W.is_contiguous() and W.dim() == 2:
+            return dequantize_fp8(W.t(), scale_inv).t()
+        return dequantize_fp8(W, scale_inv)
+
     # Get the target device from input tensor W
     target_device = W.device
 

src/axolotl/loaders/adapter.py

@@ -160,6 +160,18 @@ def load_lora(
     else:
         model = get_peft_model(model, lora_config, **model_kwargs)
 
+    # FP8 models: LoRA A/B inherit FP8 dtype from base weights, but training
+    # requires a compute dtype (bf16/fp16). Cast trainable LoRA params.
+    if cfg.torch_dtype:
+        _fp8_cast_dtype = cfg.torch_dtype
+    elif torch.cuda.is_available() and torch.cuda.is_bf16_supported():
+        _fp8_cast_dtype = torch.bfloat16
+    else:
+        _fp8_cast_dtype = torch.float16
+    for _name, param in model.named_parameters():
+        if param.requires_grad and param.dtype == torch.float8_e4m3fn:
+            param.data = param.data.to(_fp8_cast_dtype)
+
     if rank == 0:
         try:
             model.print_trainable_parameters()

src/axolotl/loaders/model.py

@@ -215,6 +215,8 @@ class ModelLoader:
             self.model_kwargs["revision"] = self.cfg.revision_of_model
         if self.cfg.use_kernels:
             self.model_kwargs["use_kernels"] = self.cfg.use_kernels
+            if "allow_all_kernels" not in self.model_kwargs:
+                self.model_kwargs["allow_all_kernels"] = self.cfg.use_kernels
         self._set_quantization_config()
         self._set_attention_config()
         self._check_model_requirements()
@@ -503,6 +505,20 @@ class ModelLoader:
         elif not is_ds_zero3:
             self.model_kwargs["device_map"] = device_map
 
+            # quantize_moe_experts quantizes expert weights on-the-fly during loading,
+            # so the actual VRAM usage is much less than bf16 estimates.
+            # When device_map is "auto", accelerate's infer_auto_device_map computes
+            # the device map at bf16 size (before quantization), causing it to offload
+            # layers to CPU, which BnB then rejects. Force single-GPU placement to
+            # prevent this. Only applies to the non-FSDP, non-ZeRO3 path (DDP/single).
+            if getattr(self.cfg, "quantize_moe_experts", False) and device_map in (
+                "auto",
+                None,
+            ):
+                self.model_kwargs["device_map"] = {
+                    "": int(os.environ.get("LOCAL_RANK", 0))
+                }
+
             cur_device = get_device_type()
             if "mps" in str(cur_device):
                 self.model_kwargs["device_map"] = "mps:0"
@@ -829,8 +845,9 @@ class ModelLoader:
     def _set_z3_leaf_modules(self):
         from deepspeed.utils import set_z3_leaf_modules
 
-        if self.cfg.model_config_type in MOE_ARCH_BLOCK:
-            moe_blocks = MOE_ARCH_BLOCK[self.cfg.model_config_type]
+        moe_type = self.cfg.model_config_type_text or self.cfg.model_config_type
+        if moe_type in MOE_ARCH_BLOCK:
+            moe_blocks = MOE_ARCH_BLOCK[moe_type]
             moe_blocks = [moe_blocks] if isinstance(moe_blocks, str) else moe_blocks
             set_z3_leaf_modules(
                 self.model,

src/axolotl/loaders/patch_manager.py

@@ -93,11 +93,13 @@ class PatchManager:
 
     def apply_pre_model_load_patches(self):
         """Apply pre-model load patches based on config."""
+        self._deactivate_hf_async_load()
         self._apply_transformers_patches()
         # self._apply_flex_attention_patches()
         self._apply_flash_attention_patches()
         self._apply_chunked_cross_entropy_patch()
         self._apply_sageattn_patches()
+        self._apply_flash_attn_4_patches()
         self._apply_fsdp_patches()
         self._apply_adapter_patches()
         self._apply_model_specific_patches()
@@ -114,6 +116,8 @@ class PatchManager:
         self._apply_patch_deepspeed_zero3()
         self._apply_voxtral_patches()
         self._apply_apertus_patches()
+        self._apply_trl_vllm_patches()
+        self._apply_trl_trainer_utils_patches()
 
     def apply_post_plugin_pre_model_load_patches(self):
         """Apply post plugin-pre_model_load load patches based on config."""
@@ -227,6 +231,15 @@ class PatchManager:
 
             patch_sageattn()
 
+    def _apply_flash_attn_4_patches(self):
+        """Auto-apply FA4 when flash_attention is enabled and FA4 is available on SM90+."""
+        if not self.cfg.flash_attention:
+            return
+
+        from axolotl.monkeypatch.attention.flash_attn_4 import patch_flash_attn_4
+
+        patch_flash_attn_4(self.model_config)
+
     def _apply_model_specific_patches(self):
         """Apply patches specific to model architectures."""
         if (
@@ -409,17 +422,27 @@ class PatchManager:
             if self.cfg.load_in_8bit:
                 apply_linear8bitlt_save_patch()
 
+    def _deactivate_hf_async_load(self):
+        """Load weights synchronously so they can be converted and not OOM."""
+        if self.cfg.load_in_4bit or self.cfg.load_in_8bit:
+            os.environ["HF_DEACTIVATE_ASYNC_LOAD"] = "1"
+
     def _apply_moe_expert_quantization_patch(self):
-        """Patch transformers weight loading to quantize MoE expert params on-the-fly."""
-        if not self.cfg.quantize_moe_experts:
+        """Patch transformers weight loading and PEFT for MoE expert quantization."""
+        has_target_params = bool(getattr(self.cfg, "lora_target_parameters", None))
+
+        if not self.cfg.quantize_moe_experts and not has_target_params:
             return
 
         from axolotl.monkeypatch.moe_quant import (
-            patch_moe_quantization_on_load,
             patch_peft_target_parameters_matching,
         )
 
-        patch_moe_quantization_on_load(self.cfg)
+        if self.cfg.quantize_moe_experts:
+            from axolotl.monkeypatch.moe_quant import patch_moe_quantization_on_load
+
+            patch_moe_quantization_on_load(self.cfg)
+
         patch_peft_target_parameters_matching()
 
     def _finalize_moe_expert_quantization(self, model: PreTrainedModel):
@@ -548,15 +571,6 @@ class PatchManager:
         LOG.info("Patching with xformers attention...")
         hijack_llama_attention()
 
-    def _patch_llama_sample_packing(self):
-        """Apply sample packing patches for LLaMA models."""
-        from axolotl.monkeypatch.llama_patch_multipack import (
-            hijack_llama_prepare_4d_mask,
-        )
-
-        LOG.info("Patching llama _prepare_4d_causal_attention_mask*...")
-        hijack_llama_prepare_4d_mask()
-
     def _patch_llama_derived_model(self):
         """Modify all llama derived models in one block."""
         if self.cfg.is_llama_derived_model and not (
@@ -568,8 +582,6 @@ class PatchManager:
                 self._patch_llama_flash_attention()
             elif self.cfg.xformers_attention:
                 self._patch_llama_xformers_attention()
-            elif self.cfg.sample_packing:
-                self._patch_llama_sample_packing()
             elif self.cfg.s2_attention:
                 raise NotImplementedError(
                     "Shifted-sparse attention not currently implemented without flash attention."
@@ -646,6 +658,50 @@ class PatchManager:
 
             patch_apertus_xielu_activation()
 
+    def _apply_trl_vllm_patches(self):
+        """Apply TRL vLLM patches for batched weight sync, NaN logprobs fix, and scalar handling."""
+        if (
+            self.cfg.rl
+            and getattr(self.cfg, "trl", None)
+            and getattr(self.cfg.trl, "use_vllm", False)
+        ):
+            from axolotl.monkeypatch.trainer.trl_vllm import patch_trl_vllm
+
+            patch_trl_vllm()
+
+    def _apply_trl_trainer_utils_patches(self):
+        """Replace trl.trainer.utils.{selective_log_softmax, entropy_from_logits} with Triton kernels."""
+        if not self.cfg.rl:
+            return
+
+        try:
+            from axolotl.monkeypatch.trainer.utils import (
+                entropy_from_logits,
+                selective_log_softmax,
+            )
+        except (ImportError, ModuleNotFoundError):
+            LOG.warning("Triton not available — skipping trl.trainer.utils patches")
+            return
+
+        import trl.trainer.utils
+
+        # Guard against repeated calls: only stash the original if trl still
+        # points at its own implementation (not our wrapper).
+        if trl.trainer.utils.selective_log_softmax is not selective_log_softmax:
+            from axolotl.monkeypatch.trainer import utils as _axolotl_trainer_utils
+
+            _axolotl_trainer_utils.selective_log_softmax_original = (
+                trl.trainer.utils.selective_log_softmax
+            )
+            trl.trainer.utils.selective_log_softmax = selective_log_softmax
+
+        if trl.trainer.utils.entropy_from_logits is not entropy_from_logits:
+            trl.trainer.utils.entropy_from_logits = entropy_from_logits
+
+        LOG.info(
+            "Patched trl.trainer.utils with Triton selective_log_softmax and entropy_from_logits"
+        )
+
     def _apply_scaling_softmax_patch(self, model: PreTrainedModel):
         """Apply Scaling Softmax (SSMax) patch.  Ref: https://arxiv.org/abs/2501.19399"""
         if self.cfg.scaling_softmax:

src/axolotl/loaders/processor.py

@@ -55,12 +55,12 @@ def load_processor(cfg: DictDefault, tokenizer: PreTrainedTokenizerBase):
         )
 
     processor_kwargs["trust_remote_code"] = cfg.trust_remote_code or False
-    processor_kwargs["tokenizer"] = tokenizer
 
     processor = processor_cls.from_pretrained(
         cfg.processor_config,
         **processor_kwargs,
     )
+    processor.tokenizer = tokenizer
 
     # Attempt to load image size from processor if available
     if (

src/axolotl/loaders/tokenizer.py

@@ -221,6 +221,14 @@ def load_tokenizer(cfg: DictDefault) -> PreTrainedTokenizer:
             if getattr(tokenizer, attr_name) is None:
                 setattr(tokenizer, attr_name, "<|endoftext|>")
 
+    # Generic fallback: if tokenizer still has no pad_token, use eos_token
+    if tokenizer.pad_token is None and tokenizer.eos_token is not None:
+        tokenizer.pad_token = tokenizer.eos_token
+        LOG.warning(
+            "Tokenizer does not have a pad_token, falling back to eos_token: %s",
+            tokenizer.eos_token,
+        )
+
     additional_special_tokens = None
     if cfg.special_tokens:
         special_tokens = cfg.special_tokens.to_dict()

src/axolotl/monkeypatch/accelerate/parallelism_config.py

@@ -78,30 +78,21 @@ def patch_parallelism_config():
 
 
 def patch_prepare_cp():
-    import functools
+    import contextlib
 
-    import torch
     from accelerate import Accelerator
 
     def patched_prepare_cp(self, *args):
         if self.parallelism_config.cp_backend == "deepspeed":
             return args
 
-        from accelerate.big_modeling import _attach_context_parallel_hooks
-        from torch.distributed.tensor.experimental import context_parallel
-        from torch.distributed.tensor.experimental._attention import set_rotate_method
-
-        cp_comm_strategy = self.parallelism_config.cp_handler.cp_comm_strategy
-        set_rotate_method(cp_comm_strategy)
-
-        self._cp_context = functools.partial(
-            context_parallel, mesh=self.torch_device_mesh["cp"]
-        )
-
-        for arg in args:
-            if isinstance(arg, torch.nn.Module):
-                _attach_context_parallel_hooks(arg)
+        @contextlib.contextmanager
+        def _noop_cp_context(
+            buffers=None, buffer_seq_dims=None, no_restore_buffers=None
+        ):
+            yield
 
+        self._cp_context = _noop_cp_context
         return args
 
     Accelerator._prepare_cp = patched_prepare_cp

src/axolotl/monkeypatch/attention/flash_attn_4.py

@@ -0,0 +1,104 @@
+"""Transparently upgrade FA2 to FA4 when available on SM90+ hardware."""
+
+import torch
+
+from axolotl.utils.logging import get_logger
+
+LOG = get_logger(__name__)
+
+
+def _get_head_dims(model_config):
+    """Extract (head_dim, head_dim_v) from a model config.
+
+    Handles composite models (e.g. Qwen3.5 VL) via text_config and
+    MLA models (DeepSeek/Kimi) that have separate Q/V head dimensions.
+    """
+    cfg = model_config
+    if hasattr(cfg, "text_config"):
+        cfg = cfg.text_config
+
+    # MLA models: Q head_dim = qk_nope + qk_rope, V head_dim = v_head_dim
+    if hasattr(cfg, "qk_nope_head_dim") and hasattr(cfg, "qk_rope_head_dim"):
+        head_dim = cfg.qk_nope_head_dim + cfg.qk_rope_head_dim
+        head_dim_v = getattr(cfg, "v_head_dim", head_dim)
+        return head_dim, head_dim_v
+
+    # Standard models
+    if hasattr(cfg, "head_dim"):
+        return cfg.head_dim, cfg.head_dim
+    if hasattr(cfg, "hidden_size") and hasattr(cfg, "num_attention_heads"):
+        head_dim = cfg.hidden_size // cfg.num_attention_heads
+        return head_dim, head_dim
+
+    return None, None
+
+
+def patch_flash_attn_4(model_config=None):
+    """Patch _lazy_imports to redirect FA2 imports to FA4 if available on supported hardware."""
+    if not torch.cuda.is_available():
+        return
+
+    major, _ = torch.cuda.get_device_capability()
+    # Matches flash_attn/cute/interface.py: arch / 10 in [9, 10, 11]
+    if major not in (9, 10, 11):
+        return
+
+    try:
+        from flash_attn.cute import (  # noqa: F401
+            flash_attn_func,
+            flash_attn_varlen_func,
+        )
+    except ImportError:
+        LOG.info(
+            "Flash Attention 4 is available for your GPU and offers faster training speeds. "
+            "To enable: pip install flash-attn-4"
+        )
+        return
+
+    # Validate head dimensions against FA4's own constraints
+    head_dim = None
+    if model_config is not None:
+        head_dim, head_dim_v = _get_head_dims(model_config)
+        if head_dim is not None:
+            try:
+                from flash_attn.cute.interface import _validate_head_dims
+            except ImportError:
+                LOG.warning(
+                    "Could not import _validate_head_dims from flash_attn.cute.interface, "
+                    "unable to verify head dimension compatibility, falling back to FA2"
+                )
+                return
+
+            # alignment = 16 // element_size; bf16/fp16 = 2 bytes -> alignment = 8
+            alignment = 8
+            try:
+                _validate_head_dims(head_dim, head_dim_v, major, alignment)
+            except AssertionError as exc:
+                LOG.warning(
+                    "Model head dimensions not supported by FA4, "
+                    "falling back to FA2: %s",
+                    exc,
+                )
+                return
+
+    import transformers.modeling_flash_attention_utils as fa_utils
+
+    if getattr(fa_utils._lazy_imports, "_axolotl_patched", False):
+        return
+
+    def _patched_lazy_imports(
+        implementation, attention_wrapper=None, allow_all_kernels=False
+    ):
+        return (
+            flash_attn_func,
+            flash_attn_varlen_func,
+            fa_utils._pad_input,
+            fa_utils._unpad_input,
+        )
+
+    _patched_lazy_imports._axolotl_patched = True
+    fa_utils._lazy_imports = _patched_lazy_imports
+    LOG.info(
+        "Flash Attention 4 enabled (head_dim=%s)",
+        head_dim if model_config else "unknown",
+    )

src/axolotl/monkeypatch/attention/flex_attn.py

@@ -64,15 +64,12 @@ def patch_flex_wrapper(**flex_attn_compile_kwargs):
                     LOG.info(
                         "Compiling flex attention with kwargs: %s. This may take a while...",
                         flex_attn_compile_kwargs,
-                        main_process_only=True,
                     )
                     self._compiled_flex_attention = torch.compile(
                         flex_attention,
                         **flex_attn_compile_kwargs,
                     )
-                    LOG.info(
-                        "Flex attention compiled successfully.", main_process_only=True
-                    )
+                    LOG.info("Flex attention compiled successfully.")
 
                 self._is_flex_compiled = True
 

src/axolotl/monkeypatch/llama_expand_mask.py

@@ -1,24 +0,0 @@
-"""
-expands the binary attention mask per 3.2.2 of https://arxiv.org/pdf/2107.02027.pdf
-"""
-
-from typing import Optional
-
-import torch
-
-from axolotl.monkeypatch.utils import mask_2d_to_4d
-
-
-def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
-    masked_zero_one_mask = mask_2d_to_4d(mask, dtype, tgt_len)
-    inverted_mask = 1.0 - masked_zero_one_mask
-
-    return inverted_mask.masked_fill(
-        inverted_mask.to(torch.bool), torch.finfo(dtype).min
-    )
-
-
-def hijack_expand_mask():
-    import transformers
-
-    transformers.models.llama.modeling_llama._expand_mask = _expand_mask

src/axolotl/monkeypatch/llama_patch_multipack.py

@@ -1,26 +0,0 @@
-"""
-Patched LlamaAttention to use torch.nn.functional.scaled_dot_product_attention
-"""
-
-from axolotl.monkeypatch.utils import (
-    patched_prepare_4d_causal_attention_mask,
-    patched_prepare_4d_causal_attention_mask_for_sdpa,
-)
-
-
-def hijack_llama_prepare_4d_mask():
-    from transformers import modeling_attn_mask_utils
-    from transformers.models.llama import modeling_llama
-
-    modeling_llama._prepare_4d_causal_attention_mask_for_sdpa = (
-        patched_prepare_4d_causal_attention_mask_for_sdpa
-    )
-    modeling_attn_mask_utils._prepare_4d_causal_attention_mask_for_sdpa = (
-        patched_prepare_4d_causal_attention_mask_for_sdpa
-    )
-    modeling_llama._prepare_4d_causal_attention_mask = (
-        patched_prepare_4d_causal_attention_mask
-    )
-    modeling_attn_mask_utils._prepare_4d_causal_attention_mask = (
-        patched_prepare_4d_causal_attention_mask
-    )

src/axolotl/monkeypatch/lora_kernels.py

@@ -51,6 +51,29 @@ QKV_PATCHES = [
     value_states = value_states.view(hidden_shape).transpose(1, 2)
 """.lstrip("\n"),
     ),
+    (
+        """
+    query_states, gate = torch.chunk(
+        self.q_proj(hidden_states).view(*input_shape, -1, self.head_dim * 2), 2, dim=-1
+    )
+    gate = gate.reshape(*input_shape, -1)
+
+    query_states = self.q_norm(query_states.view(hidden_shape)).transpose(1, 2)
+    key_states = self.k_norm(self.k_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
+    value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+""".lstrip("\n"),
+        """
+    query_states, key_states, value_states = self.apply_qkv(hidden_states)
+    query_states, gate = torch.chunk(
+        query_states.view(*input_shape, -1, self.head_dim * 2), 2, dim=-1
+    )
+    gate = gate.reshape(*input_shape, -1)
+
+    query_states = self.q_norm(query_states.view(hidden_shape)).transpose(1, 2)
+    key_states = self.k_norm(key_states.view(hidden_shape)).transpose(1, 2)
+    value_states = value_states.view(hidden_shape).transpose(1, 2)
+""".lstrip("\n"),
+    ),
 ]
 
 ORIGINAL_O_CODE = """
@@ -299,6 +322,8 @@ def get_layers(model: PeftModelForCausalLM) -> list[nn.Module]:
     if hasattr(pretrained_model, "language_model"):
         return pretrained_model.language_model.layers
     if hasattr(pretrained_model, "model"):
+        if hasattr(pretrained_model.model, "language_model"):
+            return pretrained_model.model.language_model.layers
         return pretrained_model.model.layers
 
     raise NotImplementedError(

src/axolotl/monkeypatch/moe_quant.py

@@ -1,13 +1,4 @@
-"""
-Loading-time quantization for MoE expert weights stored as 3D nn.Parameter tensors.
-
-In transformers v5, MoE models store expert weights as fused 3D tensors that BnB
-skips (only targets nn.Linear). This module patches weight loading to quantize them
-on-the-fly (4-bit via bitsandbytes parametrize, 8-bit via custom int8 parametrization),
-reducing peak VRAM from "all experts in bf16" to "one expert at a time."
-"""
-
-import os
+"""Loading-time quantization for MoE expert weights stored as 3D nn.Parameter tensors."""
 
 import bitsandbytes as bnb
 import torch
@@ -17,18 +8,20 @@ from axolotl.utils.logging import get_logger
 
 LOG = get_logger(__name__)
 
-# Module-level state for the loading-time quantization patch.
 _moe_load_state = {
     "count": 0,
     "mode": "4bit",
     "quant_type": "nf4",
     "compress_statistics": True,
     "patched": False,
+    # Module path → param names in definition order, captured before quantization.
+    # Without this, alphabetical loading order would mismatch merge order.
+    "expert_param_order": {},
 }
 
 
 class Bnb8bitParametrization(torch.nn.Module):
-    """Parametrization that dequantizes int8 row-wise quantized data on access."""
+    """Dequantizes int8 row-wise quantized data on access."""
 
     def __init__(self, row_stats: torch.Tensor):
         super().__init__()
@@ -36,7 +29,7 @@ class Bnb8bitParametrization(torch.nn.Module):
 
     @torch.no_grad()
     def forward(self, quantized_param: torch.Tensor) -> torch.Tensor:
-        # Flatten 3D+ to 2D for BnB's dequant, then reshape back.
+        """Flatten 3D+ to 2D for BnB's dequant, then reshape back."""
         orig_shape = quantized_param.shape
         if quantized_param.ndim > 2:
             quantized_param = quantized_param.reshape(-1, orig_shape[-1])
@@ -76,14 +69,11 @@ def replace_parameter_8bit(module, param_name):
 
 
 def patch_moe_quantization_on_load(cfg):
-    """Patch transformers' weight loading to quantize MoE expert params on-the-fly.
-
-    Wraps ``set_param_for_module`` so that 3D+ CUDA tensors with "expert" in their
-    name are quantized (4-bit or 8-bit) as they're loaded, keeping peak VRAM low.
-    """
+    """Patch transformers' weight loading to quantize MoE expert params on-the-fly."""
     mode = "8bit" if getattr(cfg, "load_in_8bit", False) else "4bit"
     _moe_load_state["mode"] = mode
     _moe_load_state["count"] = 0
+    _moe_load_state["expert_param_order"] = {}
 
     if _moe_load_state["patched"]:
         LOG.debug("MoE loading-time quantization patch already active")
@@ -103,14 +93,6 @@ def patch_moe_quantization_on_load(cfg):
         _moe_load_state["quant_type"] = quant_type
         _moe_load_state["compress_statistics"] = compress_statistics
 
-    # Disable async tensor loading.  Transformers' convert_and_load_state_dict_in_model
-    # uses a ThreadPoolExecutor to materialise tensors (move from safetensors → CUDA)
-    # ahead of time.  With MoE models this pre-fetches many large bf16 expert tensors
-    # onto the GPU simultaneously — long before our set_param_for_module patch can
-    # quantise and free them one-by-one — causing OOM even at <5 % of weights loaded.
-    # Sequential loading ensures only ONE bf16 expert tensor is on-GPU at a time.
-    os.environ["HF_DEACTIVATE_ASYNC_LOAD"] = "1"
-
     # Disable caching_allocator_warmup — it pre-allocates a huge tensor at bf16
     # size for all params, defeating our on-load quantization VRAM savings.
     def _noop_warmup(*args, **kwargs):
@@ -123,7 +105,6 @@ def patch_moe_quantization_on_load(cfg):
     def _patched_set_param_for_module(model, target_name, param_value, *args, **kwargs):
         original_set_param(model, target_name, param_value, *args, **kwargs)
 
-        # Quantize 3D+ expert params that BnB skipped (only on CUDA).
         if param_value.ndim >= 3 and param_value.is_cuda:
             mod_path, _, pname = target_name.rpartition(".")
             mod = model.get_submodule(mod_path) if mod_path else model
@@ -136,6 +117,13 @@ def patch_moe_quantization_on_load(cfg):
                     )
                     return
 
+                # Record definition order before parametrizations override it
+                # with alphabetical order.
+                if mod_path not in _moe_load_state["expert_param_order"]:
+                    _moe_load_state["expert_param_order"][mod_path] = list(
+                        mod._parameters.keys()
+                    )
+
                 if _moe_load_state["mode"] == "4bit":
                     replace_parameter_4bit(
                         mod,
@@ -161,20 +149,28 @@ def get_moe_quantized_count():
 
 
 def patch_peft_target_parameters_matching():
-    """Fix PEFT's _inject_parameters to use suffix matching for parametrized modules."""
+    """Fix PEFT's _inject_parameters for target_parameters on quantized MoE experts.
+
+    1. Expands short suffixes to full module paths for parametrized modules.
+    2. Iterates params in definition order (not alphabetical order) so saved
+       adapters are compatible with standard PEFT, vLLM, etc.
+    """
     if getattr(patch_peft_target_parameters_matching, "_axolotl_patched", False):
         return
-    from peft.tuners.tuners_utils import BaseTuner
 
-    original_inject = BaseTuner._inject_parameters
+    from contextlib import nullcontext
+
+    from peft.tuners.tuners_utils import BaseTuner, BaseTunerLayer
+    from peft.utils.integrations import init_empty_weights
+    from peft.utils.other import _get_submodules
 
     def _patched_inject_parameters(
         self, peft_config, model, adapter_name, low_cpu_mem_usage
     ):
-        # Patch target_parameters to use full paths for parametrized modules
         original_targets = list(peft_config.target_parameters)
         expanded = set(original_targets)
 
+        # Expand short suffixes to full paths for parametrized modules.
         for module_name, module in model.named_modules():
             if not hasattr(module, "parametrizations"):
                 continue
@@ -185,14 +181,74 @@ def patch_peft_target_parameters_matching():
                 ) and hasattr(module, param_name):
                     expanded.add(f"{module_name}.{param_name}")
 
-        peft_config.target_parameters = sorted(expanded)
-        try:
-            return original_inject(
-                self, peft_config, model, adapter_name, low_cpu_mem_usage
-            )
-        finally:
-            peft_config.target_parameters = original_targets
+        target_names_set = expanded
+
+        def strip_base_layer_from_name(module_name):
+            name = ".base_layer"
+            while name in module_name:
+                prefix, _, suffix = module_name.rpartition(name)
+                module_name = prefix + suffix
+            return module_name
+
+        def create_and_replace_param(module_name, key, param_name):
+            parent, target, target_name = _get_submodules(model, module_name)
+            unwrapped_module_name = strip_base_layer_from_name(module_name)
+            unwrapped_module = model.get_submodule(unwrapped_module_name)
+            if (
+                isinstance(unwrapped_module, BaseTunerLayer)
+                and unwrapped_module.__class__.__name__ != "ParamWrapper"
+            ):
+                raise ValueError(
+                    f"Trying to wrap an `nn.Parameter` of layer "
+                    f"'{unwrapped_module_name}' of type "
+                    f"{type(target).__name__}, which is not a valid target. "
+                    f"Make sure that this layer is not also targeted with "
+                    f"`target_modules`."
+                )
+            self._check_target_module_compatiblity(peft_config, model, target_name)
+            ctx = init_empty_weights if low_cpu_mem_usage else nullcontext
+            with ctx():
+                self._create_and_replace(
+                    peft_config,
+                    adapter_name,
+                    target,
+                    target_name,
+                    parent,
+                    current_key=key,
+                    parameter_name=param_name.rpartition(".")[-1],
+                )
+
+        # Use definition order (not alphabetical order) for parametrized modules
+        # so ParamWrapper nesting matches vanilla PEFT on a plain model.
+        expert_param_order = _moe_load_state.get("expert_param_order", {})
+
+        for module_name, module in model.named_modules():
+            if hasattr(module, "parametrizations"):
+                stored_order = expert_param_order.get(module_name)
+                if stored_order is not None:
+                    params_iter = [
+                        p for p in stored_order if p in module.parametrizations
+                    ]
+                else:
+                    # Fallback for paths that bypass model loading (e.g. unit tests).
+                    params_iter = list(module.parametrizations.keys())
+                for param_name in params_iter:
+                    key = f"{module_name}.{param_name}"
+                    if (key in target_names_set) or any(
+                        key.endswith(f".{t}") for t in target_names_set
+                    ):
+                        create_and_replace_param(module_name, key, param_name)
+                        self.targeted_parameter_names.append(key)
+            else:
+                unwrapped_module_name = strip_base_layer_from_name(module_name)
+                for param_name, _ in module.named_parameters(recurse=False):
+                    key = f"{unwrapped_module_name}.{param_name}"
+                    if (key in target_names_set) or any(
+                        key.endswith(f".{t}") for t in target_names_set
+                    ):
+                        create_and_replace_param(module_name, key, param_name)
+                        self.targeted_parameter_names.append(key)
 
     BaseTuner._inject_parameters = _patched_inject_parameters
     patch_peft_target_parameters_matching._axolotl_patched = True
-    LOG.info("Patched PEFT _inject_parameters for parametrized module suffix matching")
+    LOG.info("Patched PEFT _inject_parameters for consistent ParamWrapper ordering")

src/axolotl/monkeypatch/multipack.py

@@ -57,7 +57,9 @@ SUPPORTED_MULTIPACK_MODEL_TYPES = [
     "olmo3",
     "ministral",
     "ministral3",
+    "mistral4",
     "afmoe",
+    "nemotron",
 ]
 
 

src/axolotl/monkeypatch/ring_attn/patch.py

@@ -154,7 +154,6 @@ def register_ring_attn_from_device_mesh(
     LOG.info(
         f"Enabling ring attention sequence parallelism using DeviceMesh "
         f"dimension '{context_parallel_dim}'",
-        main_process_only=True,
     )
 
     # Extract the sequence parallel submesh

src/axolotl/monkeypatch/tiled_mlp/patch.py

@@ -85,7 +85,6 @@ def patch_tiled_mlp(model_type, use_original_mlp=True, cfg_num_shards=None):
         mlp_cls._tiled_mlp_dist_impl = None
         LOG.info(
             f"Successfully monkey-patched TiledMLP for model_type: {model_type}",
-            main_process_only=True,
         )
     except (ImportError, AttributeError) as e:
         raise RuntimeError(