c50c4acbf4
* EBFT wip * fixes * more fixeS * add missing strided module * ebft fixes for multi-turn * make ebft work with async * add example for ebft w qwen3.5 * fix for split thinking and update yaml for lora over linear attention only * enforce_eager for vllm arg in schema * fix sync weights * fix multi-gpu * handle updated sig for mm * ddp fixes * improve multi-gpu handling, don't calculate logits, adaptive completion length * chore: lint * chore: lint * support completion_mean * Address corereview feedback * clamp min IS ratio * Address PR code review * more fixes identified * address code review * Fix property from rebase conflict
87 lines
2.1 KiB
YAML
87 lines
2.1 KiB
YAML
# EBFT Structured Mode: Qwen3.5-4B (hybrid linear attention)
|
|
#
|
|
# Qwen3.5 uses hybrid attention: linear attention (conv1d) on 3/4 of layers,
|
|
# full attention every 4th layer. This tests EBFT compatibility.
|
|
#
|
|
# Prerequisites:
|
|
# 1. Start vLLM on GPU 0:
|
|
# CUDA_VISIBLE_DEVICES=0 axolotl vllm-serve examples/ebft/qwen35-4b-ebft-structured-async.yaml
|
|
#
|
|
# 2. Run training on GPU 1:
|
|
# CUDA_VISIBLE_DEVICES=1 PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
|
|
# axolotl train examples/ebft/qwen35-4b-ebft-structured-async.yaml
|
|
|
|
base_model: Qwen/Qwen3.5-4B
|
|
|
|
rl: ebft
|
|
|
|
ebft:
|
|
feature_layers: [0.25, 0.5, 0.75]
|
|
embed_method: last_token
|
|
use_whitening: false
|
|
alignment_coef: 1.0
|
|
diversity_coef: 1.0
|
|
ce_coef: 0.0
|
|
|
|
trl:
|
|
num_generations: 4
|
|
max_completion_length: 256
|
|
temperature: 0.7
|
|
use_vllm: true
|
|
vllm_server_host: 0.0.0.0
|
|
vllm_server_port: 8000
|
|
scale_rewards: true
|
|
loss_type: grpo
|
|
epsilon: 0.2
|
|
generation_kwargs:
|
|
stop_token_ids: [248044, 248046] # <|endoftext|>, <|im_end|>
|
|
chat_template_kwargs:
|
|
enable_thinking: false
|
|
async_prefetch: true
|
|
vllm_server_timeout: 300
|
|
|
|
vllm:
|
|
gpu_memory_utilization: 0.5
|
|
max_model_len: 2048
|
|
serve_module: axolotl.scripts.vllm_serve_lora
|
|
enforce_eager: true
|
|
|
|
datasets:
|
|
- path: nvidia/OpenCodeInstruct
|
|
type: ebft_opencode.transform
|
|
split: train[:500]
|
|
|
|
sequence_len: 1024
|
|
micro_batch_size: 1
|
|
gradient_accumulation_steps: 4
|
|
num_epochs: 1
|
|
max_steps: 10
|
|
|
|
learning_rate: 5.0e-6
|
|
optimizer: adamw_torch_fused
|
|
lr_scheduler: cosine
|
|
warmup_steps: 3
|
|
weight_decay: 0.01
|
|
|
|
adapter: lora
|
|
lora_r: 16
|
|
lora_alpha: 32
|
|
lora_dropout: 0.0
|
|
# Target full-attention q/k/v/o on layers 3,7,11,15,19,23,27,31 + MLP on all layers
|
|
# Avoids linear_attn modules (in_proj_qkv, in_proj_z, etc.) which break vLLM LoRA
|
|
lora_target_modules: ".*\\.layers\\.(3|7|11|15|19|23|27|31)\\.self_attn\\.(q|k|v|o)_proj|.*\\.mlp\\.(gate|up|down)_proj"
|
|
|
|
bf16: auto
|
|
flash_attention: true
|
|
gradient_checkpointing: true
|
|
|
|
special_tokens:
|
|
pad_token: "<|endoftext|>"
|
|
|
|
val_set_size: 0.0
|
|
output_dir: ./outputs/ebft-qwen35-4b-structured-async
|
|
|
|
wandb_project: ebft
|
|
logging_steps: 1
|
|
save_steps: 50
|