rename to specify fsdp

cicd/multigpu.py

@@ -68,7 +68,7 @@ def run_cmd(cmd: str, run_folder: str):
 @app.function(
     image=cicd_image,
     gpu=GPU_CONFIG,
-    timeout=90 * 60,
+    timeout=60 * 60,
     cpu=8.0,
     memory=131072 * N_GPUS,
     volumes=VOLUME_CONFIG,

examples/llama-4/README.md

@@ -7,10 +7,4 @@
 - [Text Single GPU (H100) QLoRA](./scout-qlora-single-h100.yaml)
 - [Text Multi GPU QLoRA w/ FSDP1](./scout-qlora-fsdp1.yaml)
 
-Our Single H100 implementation for Llama 4 Scout uses only 68.5GB VRAM for post-training with 4k context length @ 546 tokens/second. [WandB logs here](https://wandb.ai/axolotl-ai/llama4-sft/runs/zic56rhd)
-
-### Llama 4 Maverick 17Bx128Experts (400B)
-
-- [Text Multi GPU QLoRA w/FSDP1](./maverick-qlora-fsdp1.yaml)
-
-Our 4xH100 implementation for Llama 4 Maverick uses 79.5GB VRAM/GPU for post-training with 4k context length @ 206 tokens/second. [WandB logs here.](https://wandb.ai/axolotl-ai/llama-sft/runs/siyvwuxc?nw=nwuserwinglian)
+Our Single GPU implementation for Llama 4 Scout uses only 68.5GB VRAM for post-training with 4k context length @ 546 tokens/second.

examples/llama-4/maverick-qlora-fsdp1.yaml

@@ -1,89 +0,0 @@
-base_model: axolotl-quants/Llama-4-Maverick-17B-128E-Linearized-bnb-nf4-bf16
-model_type: Llama4ForConditionalGeneration
-# Automatically upload checkpoint and final model to HF
-# hub_model_id: username/custom_model_name
-
-strict: false
-
-plugins:
-  - axolotl.integrations.liger.LigerPlugin
-
-liger_glu_activation: true
-liger_rms_norm: true
-liger_layer_norm: true
-
-llama4_linearized_experts: true
-load_in_4bit: true
-adapter: qlora
-lora_r: 32
-lora_alpha: 64
-lora_target_modules:
-  - self_attn.q_proj
-  - self_attn.k_proj
-  - self_attn.v_proj
-  - self_attn.o_proj
-  - shared_expert.gate_proj
-  - shared_expert.up_proj
-  - shared_expert.down_proj
-  # - experts.gate_projs.[0-9]+$
-  # - experts.up_projs.[0-9]+$
-  # - experts.down_projs.[0-9]+$
-lora_modules_to_save:
-# - lm_head
-# - embed_tokens
-
-chat_template: llama4
-datasets:
-  - path: mlabonne/FineTome-100k
-    type: chat_template
-    split: train[:20%]
-    field_messages: conversations
-    message_property_mappings:
-      role: from
-      content: value
-
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.0
-output_dir: ./outputs/out
-
-sequence_len: 4096
-sample_packing: true
-pad_to_sequence_len: true
-
-gradient_accumulation_steps: 1
-micro_batch_size: 1
-num_epochs: 1
-optimizer: adamw_torch_fused
-lr_scheduler: cosine
-learning_rate: 1e-4
-
-bf16: true
-tf32: true
-
-logging_steps: 1
-flash_attention: true
-
-gradient_checkpointing: offload
-gradient_checkpointing_kwargs:
-  use_reentrant: false
-
-warmup_steps: 20
-evals_per_epoch: 1
-saves_per_epoch: 1
-weight_decay: 0.0
-fsdp:
-  - auto_wrap
-  - full_shard
-fsdp_config:
-  fsdp_transformer_layer_cls_to_wrap: Llama4TextDecoderLayer
-  fsdp_limit_all_gathers: true
-  fsdp_sync_module_states: true
-  fsdp_offload_params: true
-  fsdp_use_orig_params: false
-  fsdp_cpu_ram_efficient_loading: true
-  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
-  fsdp_state_dict_type: FULL_STATE_DICT
-  fsdp_sharding_strategy: FULL_SHARD
-special_tokens:
-  pad_token: <|finetune_right_pad_id|>
-  eos_token: <|eot|>

requirements.txt

@@ -12,7 +12,7 @@ liger-kernel==0.5.6
 packaging==23.2
 
 peft==0.15.1
-transformers==4.51.1
+transformers==4.51.0
 tokenizers>=0.21.1
 accelerate==1.6.0
 datasets==3.5.0

src/axolotl/integrations/liger/__init__.py

@@ -185,7 +185,5 @@ class LigerPlugin(BasePlugin):
                 rms_norm=cfg.liger_rms_norm,
                 layer_norm=cfg.liger_layer_norm,
             )
-        else:
-            logging.warning(
-                f"Unsupported model config type: {cfg.model_config_type}. Liger not applied."
-            )
+        elif cfg.model_config_type in ["deepseek_v3"]:
+            raise ValueError(f"Unsupported model config type: {cfg.model_config_type}")

src/axolotl/integrations/liger/models/llama4.py

@@ -3,7 +3,6 @@ Liger FLCE for llama4
 """
 
 import sys
-from copy import deepcopy
 from typing import List, Optional, Tuple, Union
 
 import torch
@@ -159,16 +158,7 @@ def apply_liger_kernel_to_llama4(
     if rms_norm:
         modeling_llama4.Llama4TextRMSNorm = LigerRMSNorm
     if glu_activation:
-
-        def _liger_swiglu_mlp_wrapper(config, intermediate_size=None, **kwargs):
-            "Accepts intermediate_size to pass to LigerSwiGLUMLP"
-            # clone config to avoid modifying the original
-            config = deepcopy(config)
-            if intermediate_size:
-                setattr(config, "intermediate_size", intermediate_size)
-            return LigerSwiGLUMLP(config, **kwargs)
-
-        modeling_llama4.Llama4TextMLP = _liger_swiglu_mlp_wrapper
+        modeling_llama4.Llama4TextMLP = LigerSwiGLUMLP
     if layer_norm:
         modeling_llama4.nn.LayerNorm = LigerLayerNorm