tocmo0nlord/axolotl
1
0
Fork 0
Code Issues Pull Requests Actions 3 Packages Projects Releases Wiki Activity

chore: cleanup post release v0.16 (#3577)

Browse Source 
* fix: remove unneeded debug log

* fix: cleanup

* feat: add dense gemma config and cleanup

* feat: add cce support

* update notes and set torch compile

* fix patch for new number of return vals

* fixes for gemma4

* fix packing bug

* use updated cce for mm

* fix: pass in kv cache func when avail for transformers 5.5

* feat: update examples with flex variant and readme

* gemma4 lora attention kernels

---------

Co-authored-by: Wing Lian <wing.lian@gmail.com>
Co-authored-by: Wing Lian <wing@axolotl.ai>
This commit is contained in:
NanoCode012
2026-04-07 00:10:52 +07:00
committed by GitHub
parent dc638e723f
commit 149178ddb7
15 changed files with 665 additions and 60 deletions
Download Patch File Download Diff File Expand all files Collapse all files

33
examples/gemma4/26b-a4b-moe-qlora.yaml
View File

@@ -1,19 +1,12 @@
# Gemma 4 26B-A4B MoE QLoRA with ScatterMoE kernels
#
# Validated: 50 steps on FineTome-100k, loss 7.4 -> 2.4, single RTX 5090 (32GB)
# Validated: 50 steps on FineTome-100k, loss 8.8 -> 1.8, single RTX 5090 (32GB)
# torch_compile=true: 21 GiB peak VRAM, ~230 tok/s, 336s total
#
# Key notes:
# - Flash Attention 2 is NOT supported (global_head_dim=512 > FA2 max of 256).
# Use sdp_attention instead.
# - Gemma 4 is multimodal (text+vision+audio). For text-only SFT, restrict
# LoRA to the text backbone via lora_target_linear_modules regex.
# - MoE experts use `experts_implementation: scattermoe` — Gemma 4 embeds MoE
# directly in the decoder layer (no SparseMoeBlock), so we register ScatterMoE
# via the transformers ExpertsInterface.
# - Expert LoRA targets are `experts.gate_up_proj` / `experts.down_proj`
# (no `mlp.` prefix, unlike Qwen/Mixtral).
# - micro_batch_size: 1 fits 2048 seq_len on 32GB GPU with SDP attention.
# Use micro_batch_size: 4 with 1024 seq_len, or on 48GB+ GPUs.
# - Max sequence length on 32GB GPU: 2048 (micro_batch_size=1, SDP attention).
# 4096 seq_len OOMs due to head_dim=512 math SDP materializing full score matrix.
# Use 48GB+ GPUs for longer sequences or multi-GPU with FSDP.
base_model: google/gemma-4-26B-A4B
@@ -24,7 +17,7 @@ plugins:
use_kernels: true
use_scattermoe: true
experts_implementation: scattermoe
torch_compile: false
torch_compile: true
liger_layer_norm: true
liger_rope: true
liger_rms_norm: true
@@ -54,12 +47,9 @@ lora_r: 16
lora_alpha: 32
lora_dropout: 0
# Restrict LoRA to text backbone only (skip vision/audio encoders).
# lora_target_modules is intentionally empty — all module targeting is done
# via regex in lora_target_linear_modules below.
lora_target_modules: []
lora_target_linear_modules:
- language_model\.model\.layers\.\d+\.self_attn\.(q|k|v|o)_proj
# Restrict LoRA to text backbone only (skip vision/audio encoders)
# using regex to match only the text decoder attention projections.
lora_target_modules: 'model.language_model.layers.[\d]+.(_checkpoint_wrapped_module.)?(mlp|self_attn).(up|down|gate|q|k|v|o)_proj'
# MoE expert LoRA (3D Parameter tensors, not nn.Linear)
lora_target_parameters:
@@ -73,7 +63,7 @@ lora_o_kernel: false
bnb_config_kwargs:
bnb_4bit_use_double_quant: true
wandb_project: gemma4-qlora
wandb_project:
wandb_entity:
wandb_watch:
wandb_name:
@@ -93,8 +83,7 @@ gradient_checkpointing: true
activation_offloading: true
logging_steps: 1
# FA2 not supported — Gemma4 global_head_dim=512 exceeds FA2 max of 256
flash_attention: false
# FA2 not supported
sdp_attention: true
warmup_ratio: 0.1