[ROCm][Quantization][4/N] refactor quark_moe fp8 w/ oracle (#43721)

Signed-off-by: Bowen Bao <bowenbao@amd.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
Bowen Bao
2026-06-22 15:58:03 -07:00
committed by GitHub
co-authored by Andreas Karatzas
parent 183b5f27ea
commit c97e8f99d6
5 changed files with 90 additions and 103 deletions
@@ -0,0 +1,6 @@
model_name: "amd/Qwen3-30B-A3B-Thinking-2507-FP8"
accuracy_threshold: 0.81
num_questions: 1319
num_fewshot: 5
max_tokens: 1024
server_args: "--max-model-len 4096 --gpu-memory-utilization 0.85"
@@ -0,0 +1,6 @@
model_name: "amd/Qwen3-30B-A3B-Thinking-2507-PTPC-FP8"
accuracy_threshold: 0.81
num_questions: 1319
num_fewshot: 5
max_tokens: 1024
server_args: "--max-model-len 4096 --gpu-memory-utilization 0.85"
@@ -3,3 +3,5 @@ Qwen2.5-VL-3B-Instruct-FP8-dynamic.yaml
Qwen1.5-MoE-W4A16-CT.yaml
DeepSeek-V2-Lite-Instruct-FP8.yaml
Qwen3-Next-FP8-EP2_MI355.yaml
Qwen3-30B-A3B-Thinking-2507-FP8.yaml
Qwen3-30B-A3B-Thinking-2507-PTPC-FP8.yaml
@@ -48,6 +48,7 @@ def run_gsm8k_eval(eval_config: dict, server_url: str) -> dict:
results = evaluate_gsm8k(
num_questions=eval_config["num_questions"],
num_shots=eval_config["num_fewshot"],
max_tokens=eval_config.get("max_tokens", 256),
host=host,
port=port,
request_timeout_seconds=request_timeout_seconds,