Specializes Qwen3.5-397B-A17B for AMD GPU kernel optimization tasks -- writing Triton kernels, debugging ROCm issues, and optimizing performance on AMD Instinct GPUs. Trained on 104 multi-turn agent trajectories from the amdpilot dataset.

Version History

Version	Train Loss	Eval Loss	Key Change	HuggingFace
v1	0.163	n/a	Baseline pipeline	v1
v2	0.085	n/a	3-view data extraction (-48% loss)	v2
v3	0.059	0.044	Recipe fix: 10x steps, 2x rank, eval (-31% loss)	this repo

Training Details

Parameter	Value
Base model	Qwen/Qwen3.5-397B-A17B (MoE, 17B active)
Hardware	8x AMD Instinct MI355X (ROCm 7.2)
LoRA rank / alpha	32 / 64
Target modules	all (13 types)
Trainable params	128.5M / 396.9B (0.032%)
Dataset	296 examples (3-view from 104 trajectories)
Cutoff length	32,768 tokens
Epochs / Steps	10 / 130
Batch size	8 (1 per device x 8 GPUs)
Learning rate	2e-5 (cosine schedule)
Weight decay	0.01
Training time	5h 10min
Framework	LLaMA-Factory + DeepSpeed ZeRO-3 + PEFT 0.18.1

Eval Loss Trajectory

Step	Epoch	Eval Loss
20	1.5	0.0618
40	3.1	0.0539
60	4.6	0.0491
80	6.2	0.0461
100	7.7	0.0446
120	9.2	0.0443
130	10.0	0.0442

Eval loss decreases monotonically with no overfitting. wandb run.

Usage

Load with PEFT

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("JinnP/Qwen3.5-397B-A17B-LoRA-SFT-v3")
model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3.5-397B-A17B", device_map="auto", torch_dtype="bfloat16"
)
model = PeftModel.from_pretrained(model, "JinnP/Qwen3.5-397B-A17B-LoRA-SFT-v3")

Serve with vLLM (LoRA hot-loading)

bash

python -m vllm.entrypoints.openai.api_server \
  --model Qwen/Qwen3.5-397B-A17B \
  --enable-lora \
  --lora-modules amdpilot=JinnP/Qwen3.5-397B-A17B-LoRA-SFT-v3 \
  --tensor-parallel-size 8

Merge with LLaMA-Factory

bash

llamafactory-cli export \
  --model_name_or_path Qwen/Qwen3.5-397B-A17B \
  --adapter_name_or_path JinnP/Qwen3.5-397B-A17B-LoRA-SFT-v3 \
  --template qwen3_5_nothink \
  --finetuning_type lora \
  --export_dir saves/qwen35-397b-merged

Dataset

JinnP/amdpilot-lora-sft-dataset -- 104 multi-turn agent trajectories:

94 KernelBench Triton kernel optimization tasks
4 SGLang/vLLM bugfix and feature tasks
4 frontier bugfix trajectories
Processed into 296 training examples using 3-view extraction (bookend + full + solution chunks)

Framework Versions

PEFT 0.18.1
Transformers 5.2.0
PyTorch 2.9.1+rocm7.2.0
Datasets 4.0.0
Tokenizers 0.22.2