[AMD] Fix garbled unquantized Qwen3-30B-A3B output on ROCm/aiter where the aiter CK fused-MoE falls back to Triton with pre-shuffled weights (#28244)

Co-authored-by: Xinyu Jiang <xinyuj2@andrew.cmu.edu>

python/sglang/srt/layers/quantization/unquant.py

@@ -238,8 +238,13 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, MultiPlatformOp):
             set_weight_attrs(w2_weight_bias, extra_weight_attrs)
 
     def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
-        _should_use_aiter_moe = _use_aiter and (
+        _should_use_aiter_moe = (
-            get_moe_runner_backend().is_auto() or get_moe_runner_backend().is_aiter()
+            _use_aiter
+            and (
+                get_moe_runner_backend().is_auto()
+                or get_moe_runner_backend().is_aiter()
+            )
+            and self._aiter_ck_moe_supported(layer)
         )
         if _should_use_aiter_moe:
             copy_or_rebind_param(
@@ -392,6 +397,11 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, MultiPlatformOp):
 
         param.data = param.data.reshape(expected_shape)
 
+    def _aiter_ck_moe_supported(self, layer) -> bool:
+        # aiter CK fused-MoE requires intermediate_size_per_partition to be 128-aligned
+        # (GemmSpec=Default; otherwise CK raises "not support this GEMM problem").
+        return layer.intermediate_size_per_partition % 128 == 0
+
     def create_moe_runner(
         self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig
     ):
@@ -410,7 +420,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, MultiPlatformOp):
             backend = MoeRunnerBackend.TRITON
         self.runner = MoeRunner(backend, moe_runner_config)
 
-        # Separate runner so CK-shape errors fall back to self.runner on every call.
+        # aiter CK fused-MoE only supports 128-aligned shapes; otherwise use triton.
         self._aiter_runner: Optional[MoeRunner] = None
         if (
             _use_aiter
@@ -420,7 +430,22 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, MultiPlatformOp):
             )
             and get_moe_a2a_backend().supports_aiter()
         ):
-            self._aiter_runner = MoeRunner(MoeRunnerBackend.AITER, moe_runner_config)
+            if self._aiter_ck_moe_supported(layer):
+                self._aiter_runner = MoeRunner(
+                    MoeRunnerBackend.AITER, moe_runner_config
+                )
+            elif get_moe_runner_backend().is_aiter():
+                raise ValueError(
+                    "moe_runner_backend=aiter is not supported for "
+                    f"intermediate_size_per_partition={layer.intermediate_size_per_partition}; "
+                    "use --moe-runner-backend triton."
+                )
+            else:
+                logger.warning_once(
+                    "aiter CK fused-MoE does not support "
+                    f"intermediate_size_per_partition={layer.intermediate_size_per_partition}; "
+                    "using triton MoE runner."
+                )
 
     @property
     def load_up_proj_weight_first(self) -> bool:
@@ -522,20 +547,12 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, MultiPlatformOp):
                     AiterMoeQuantInfo,
                 )
 
-                try:
                 quant_info = AiterMoeQuantInfo(
                     w13_weight=layer.w13_weight,
                     w2_weight=layer.w2_weight,
                     expert_mask=layer.dispatcher.expert_mask_gpu,
                 )
                 return self._aiter_runner.run(dispatch_output, quant_info)
-                except RuntimeError as e:
-                    # AITER CK fused_moe may not support all GEMM dimensions
-                    # (e.g. Gemma4 MoE with 128 experts x 704 intermediate size)
-                    logger.warning_once(
-                        f"AITER CK fused_moe failed ({e}), "
-                        "falling back to Triton MoE runner."
-                    )
 
             quant_info = TritonMoeQuantInfo(
                 w13_weight=layer.w13_weight,

test/registered/amd/accuracy/mi35x/test_qwen3_moe_eval_mi35x.py

@@ -0,0 +1,230 @@
+"""MI35x Qwen3 MoE (unquantized) GSM8K Completion Evaluation Test (8-GPU)
+
+Tests unquantized (bf16) Qwen3 MoE (Qwen/Qwen3-30B-A3B) using a few-shot GSM8K
+completion benchmark on MI35x with aiter enabled (SGLANG_USE_AITER=1).
+
+Registry: nightly-amd-8-gpu-mi35x suite
+"""
+
+import ast
+import os
+import re
+import time
+import unittest
+from dataclasses import dataclass
+from typing import List, Optional, Tuple
+
+import numpy as np
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.ci.ci_register import register_amd_ci
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    is_in_ci,
+    popen_launch_server,
+    write_github_step_summary,
+)
+from sglang.utils import download_and_cache_file, read_jsonl
+
+# Register for AMD CI - MI35x Qwen3 MoE accuracy test (~15 min)
+register_amd_ci(est_time=1200, suite="nightly-amd-8-gpu-mi35x", nightly=True)
+
+INVALID = -9999999
+
+
+@dataclass
+class ModelConfig:
+    """Configuration for a model to test."""
+
+    model_path: str
+    tp_size: int = 8
+    accuracy_threshold: float = 0.50
+    other_args: Optional[List[str]] = None
+    env_vars: Optional[dict] = None
+    timeout: Optional[int] = None
+
+    def __post_init__(self):
+        if self.other_args is None:
+            self.other_args = []
+        if self.env_vars is None:
+            self.env_vars = {}
+
+
+# Unquantized (bf16) Qwen3 MoE on MI35x with aiter enabled.
+MI35X_QWEN3_MOE_MODELS = [
+    ModelConfig(
+        model_path="Qwen/Qwen3-30B-A3B",
+        tp_size=8,
+        accuracy_threshold=0.75,
+        other_args=[
+            "--max-running-requests",
+            "128",
+            "--mem-fraction-static",
+            "0.8",
+            "--trust-remote-code",
+        ],
+        # ServerArgs routes the unquantized Qwen3 MoE runner to triton on ROCm with
+        # aiter (aiter CK can't run the TP-sharded intermediate, e.g. 768 // 8).
+        env_vars={"SGLANG_USE_AITER": "1"},
+    ),
+]
+
+
+def get_one_example(lines, i, include_answer):
+    """Format a single GSM8K example."""
+    ret = "Question: " + lines[i]["question"] + "\nAnswer:"
+    if include_answer:
+        ret += " " + lines[i]["answer"]
+    return ret
+
+
+def get_few_shot_examples(lines, k):
+    """Get k few-shot examples for prompting."""
+    ret = ""
+    for i in range(k):
+        ret += get_one_example(lines, i, True) + "\n\n"
+    return ret
+
+
+def get_answer_value(answer_str):
+    """Extract numerical answer from response."""
+    answer_str = answer_str.replace(",", "")
+    numbers = re.findall(r"\d+", answer_str)
+    if len(numbers) < 1:
+        return INVALID
+    try:
+        return ast.literal_eval(numbers[-1])
+    except SyntaxError:
+        return INVALID
+
+
+def run_gsm8k_benchmark(
+    base_url: str,
+    num_questions: int = 200,
+    num_shots: int = 5,
+    parallel: int = 64,
+) -> Tuple[float, float, float]:
+    """Run GSM8K few-shot completion benchmark."""
+    import sglang as sgl
+    from sglang.lang.backend.runtime_endpoint import RuntimeEndpoint
+
+    url = "https://raw.githubusercontent.com/openai/grade-school-math/master/grade_school_math/data/test.jsonl"
+    data_path = download_and_cache_file(url)
+    lines = list(read_jsonl(data_path))
+
+    few_shot_examples = get_few_shot_examples(lines, num_shots)
+
+    questions = []
+    labels = []
+    for i in range(len(lines[:num_questions])):
+        questions.append(get_one_example(lines, i, False))
+        labels.append(get_answer_value(lines[i]["answer"]))
+    assert all(l != INVALID for l in labels)
+    arguments = [{"question": q} for q in questions]
+
+    @sgl.function
+    def few_shot_gsm8k(s, question):
+        s += few_shot_examples + question
+        s += sgl.gen(
+            "answer", max_tokens=512, stop=["Question", "Assistant:", "<|separator|>"]
+        )
+
+    backend = RuntimeEndpoint(base_url)
+    sgl.set_default_backend(backend)
+
+    tic = time.perf_counter()
+    states = few_shot_gsm8k.run_batch(
+        arguments, temperature=0, num_threads=parallel, progress_bar=True
+    )
+    latency = time.perf_counter() - tic
+
+    preds = [get_answer_value(states[i]["answer"]) for i in range(len(states))]
+    acc = np.mean(np.array(preds) == np.array(labels))
+    invalid = np.mean(np.array(preds) == INVALID)
+
+    return float(acc), float(invalid), float(latency)
+
+
+class TestQwen3MoeEvalMI35x(unittest.TestCase):
+    """Unquantized Qwen3 MoE GSM8K Completion Evaluation Test for AMD MI35x."""
+
+    @classmethod
+    def setUpClass(cls):
+        cls.models = MI35X_QWEN3_MOE_MODELS
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.num_questions = int(os.environ.get("GSM8K_NUM_QUESTIONS", "200"))
+
+    def test_qwen3_moe_accuracy(self):
+        """Test unquantized Qwen3 MoE with GSM8K completion benchmark."""
+        all_results = []
+        summary = "### Qwen3 MoE Models (MI35x)\n\n"
+        summary += "| Model | TP | Accuracy | Threshold | Status |\n"
+        summary += "| ----- | -- | -------- | --------- | ------ |\n"
+
+        for config in self.models:
+            with self.subTest(model=config.model_path):
+                print(f"\n{'='*60}")
+                print(f"Testing: {config.model_path}")
+                print(f"{'='*60}")
+
+                env = os.environ.copy()
+                for key, value in config.env_vars.items():
+                    env[key] = value
+
+                other_args = list(config.other_args)
+                other_args.extend(["--tp", str(config.tp_size)])
+                timeout = config.timeout or DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH
+
+                try:
+                    process = popen_launch_server(
+                        model=config.model_path,
+                        base_url=self.base_url,
+                        timeout=timeout,
+                        other_args=other_args,
+                        env=env,
+                    )
+
+                    try:
+                        acc, invalid, latency = run_gsm8k_benchmark(
+                            self.base_url, num_questions=self.num_questions
+                        )
+                        passed = acc >= config.accuracy_threshold
+                        status = "✅ PASS" if passed else "❌ FAIL"
+                        print(
+                            f"  accuracy={acc:.3f} threshold={config.accuracy_threshold} {status}"
+                        )
+
+                        all_results.append(
+                            {
+                                "model": config.model_path,
+                                "accuracy": acc,
+                                "passed": passed,
+                            }
+                        )
+                        summary += f"| {config.model_path} | {config.tp_size} | {acc:.3f} | {config.accuracy_threshold} | {status} |\n"
+
+                    finally:
+                        kill_process_tree(process.pid)
+
+                except Exception as e:
+                    summary += f"| {config.model_path} | {config.tp_size} | N/A | {config.accuracy_threshold} | ❌ ERROR |\n"
+                    all_results.append(
+                        {
+                            "model": config.model_path,
+                            "accuracy": None,
+                            "passed": False,
+                            "error": str(e),
+                        }
+                    )
+
+        if is_in_ci():
+            write_github_step_summary(summary)
+
+        failed = [r for r in all_results if not r["passed"]]
+        if failed:
+            raise AssertionError(f"Failed models: {[r['model'] for r in failed]}")
+
+
+if __name__ == "__main__":
+    unittest.main()