e4032fc90f
* upgrade to torchao 0.17.0 * chore: lint * refactor attention handling * replace legacy attention boolean flags with capability properties Replace checks with capability-based properties derived from attn_implementation This separates three concerns that were conflated under flash_attention: 1. Backend selection -> attn_implementation enum 2. Packing capability -> attn_supports_packing property 3. Flash-attn library dependency -> attn_uses_flash_lib property * compute attn capability flags in normalizer instead of properties * make attn_implementation the single source of truth * move attention-dependent validators to mode=after * migrate remaining consumers to canonical attn_implementation * expand attention tests + rewrite docs * migrate example configs to canonical attn_implementation * update doc snippets + reject gemma4-hybrid with non-FA2 backend * remove dead gemma4 branch in _set_attention_config * fix duplicate attn_implementation in gpt-oss yamls and flaky caplog tests * drop "Phase 2" naming from attn-implementation tests * regroup attn_implementation tests by feature concern * clean up verbose comments and remove MD Signed-off-by: Wing Lian <wing@axolotl.ai> Co-authored-by: Axolotl Swarm <no-reply@axolotl.ai> * fix(collator): pass return_dict=True at apply_chat_template top level for transformers 5.x In transformers 5.x, ProcessorMixin.apply_chat_template gained its own `return_dict` parameter (defaulting to False). When return_dict=False and tokenize=True the method returns out["input_ids"] directly — a 2-D tensor — rather than the full BatchFeature dict. The old code placed `return_dict=True` inside processor_kwargs. In transformers 5.x those kwargs are forwarded to the underlying processor call self(...) where _merge_kwargs silently ignores any key not present in MllamaProcessorKwargs (emitting a warning). The outer return_dict therefore stayed False, apply_chat_template returned the raw input_ids tensor, and the subsequent `batch["input_ids"]` attempted to index a 2-D tensor with the 9-character string "input_ids", producing: IndexError: too many indices for tensor of dimension 2 The fix is to pass return_dict=True as a top-level keyword argument to apply_chat_template (where it is actually consumed) and remove it from processor_kwargs (where it was silently dropped). No version guard is needed: transformers is pinned to ==5.5.4 in pyproject.toml. Adds a unit-level regression test (tests/test_mm_chat_collator.py) that mocks the processor to return a raw tensor when apply_chat_template is called without top-level return_dict=True, verifying the four invariants: process_rows returns a dict, input_ids is 2-D, labels is 2-D, and apply_chat_template receives return_dict=True as a top-level kwarg. Fixes: tests/e2e/test_llama_vision.py::TestLlamaVision::test_lora_llama_vision_multimodal_dataset Fixes: tests/e2e/test_llama_vision.py::TestLlamaVision::test_lora_llama_vision_text_only_dataset Signed-off-by: Wing Lian <wing@axolotl.ai> Co-authored-by: Axolotl Swarm <no-reply@axolotl.ai> * fix(collator): process_rows returns dict (BatchFeature) shape Two related changes for the multimodal chat collator under transformers 5.x: 1. Wrap apply_chat_template result in dict(...) so process_rows returns a plain dict rather than a BatchFeature instance. BatchFeature is a Mapping but not a dict; downstream code that did batch["labels"] = self.processing_strategy.process_labels(batch["input_ids"]) would index on a tensor when the result wasn't dict-shaped, raising IndexError: too many indices for tensor of dimension 2 2. Soften the regression test's contract from `dict` to `Mapping` so it exercises the actual semantic guarantee (key/value access) rather than the implementation detail (dict vs BatchFeature). Test guards against the original transformers 5.x breakage where apply_chat_template's return_dict default went from True to False. Includes regression test under tests/test_mm_chat_collator.py. Bug surfaced via swarm dispatch task_01KQHPNAYD8XARSNSDJVW1GPF6 against attn-implementation-refactor; squash-merged from agent commits 4de886fd + dc9fcf4f. Signed-off-by: Wing Lian <wing@axolotl.ai> --------- Signed-off-by: Wing Lian <wing@axolotl.ai> Co-authored-by: Axolotl Swarm <no-reply@axolotl.ai>
Nemotron-H (nvidia/NVIDIA-Nemotron-3-*)
Hybrid Mamba2 / Attention / MoE architecture (model_type: nemotron_h).
| Model | Total params | Active params | Layers |
|---|---|---|---|
| NVIDIA-Nemotron-3-Super-120B-A12B-BF16 | 120B | ~12B | 88 |
| NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 | 30B | ~3B | — |
Requirements
pip install mamba-ssm causal-conv1d # fast Mamba2 CUDA kernels
Architecture notes
- Three block types per layer: Mamba2 (selective SSM), Attention (sparse), MoE (mixture-of-experts).
- Only ~12 out of 88 blocks are attention layers (120B variant).
- MLP activation is
relu2viamlp_hidden_act(not the usualhidden_act).
LoRA kernel patches
All three LoRA Triton kernel patches must be disabled:
lora_qkv_kernel: false # attention lives in NemotronHBlock.mixer, not layer.self_attn
lora_o_kernel: false # same reason
lora_mlp_kernel: false # relu2 (mlp_hidden_act) is not supported by lora_mlp_kernel
MoE expert weights
NemotronH experts store up_proj and down_proj as 3D nn.Parameter tensors
(shape [num_experts, out_dim, in_dim]), not nn.Linear modules — there is no
gate_proj. To fine-tune them alongside attention, use lora_target_parameters
instead of lora_target_modules:
lora_target_parameters:
- up_proj
- down_proj
Limitations
- MoE Triton kernels:
lora_mlp_kernelis not supported for NemotronH's MoE expert layers. The expert weights are 3Dnn.Parametertensors (notnn.Linear), which the Triton kernel does not support. Keeplora_mlp_kernel: false. - Gradient checkpointing: Only supported when
sample_packing: true. Without sample packing the upstream model markssupports_gradient_checkpointing = False.