[Bugfix] Fix DeepGemm E8M0 accuracy degradation for Qwen3.5 FP8 on Blackwell (#38083)

Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>

tests/evals/gsm8k/configs/Qwen3.5-35B-A3B-DEP2.yaml

 model_name: "Qwen/Qwen3.5-35B-A3B"
-accuracy_threshold: 0.86
+accuracy_threshold: 0.84
+tolerance: 0.03
 num_questions: 1319
 num_fewshot: 5
 server_args: >-

tests/evals/gsm8k/configs/Qwen3.5-35B-A3B-FP8-DEP2.yaml

 model_name: "Qwen/Qwen3.5-35B-A3B-FP8"
-accuracy_threshold: 0.86
+accuracy_threshold: 0.79
+tolerance: 0.03
 num_questions: 1319
 num_fewshot: 5
 server_args: >-

tests/evals/gsm8k/configs/Qwen3.5-397B-A17B-NVFP4-DEP2.yaml

+model_name: "nvidia/Qwen3.5-397B-A17B-NVFP4"
+accuracy_threshold: 0.88
+tolerance: 0.03
+num_questions: 1319
+num_fewshot: 5
+server_args: >-
+  --max-model-len 4096
+  --data-parallel-size 2
+  --enable-expert-parallel

tests/evals/gsm8k/configs/models-qwen35-blackwell.txt

 Qwen3.5-35B-A3B-DEP2.yaml
+Qwen3.5-35B-A3B-FP8-DEP2.yaml
+Qwen3.5-397B-A17B-NVFP4-DEP2.yaml
\ No newline at end of file

tests/evals/gsm8k/test_gsm8k_correctness.py

@@ -19,8 +19,6 @@ from vllm.platforms import current_platform
 
 from .gsm8k_eval import evaluate_gsm8k
 
-TOL = 0.08  # Absolute tolerance for accuracy comparison
-
 
 def run_gsm8k_eval(eval_config: dict, server_url: str) -> dict:
     """Run GSM8K evaluation using our isolated script."""
@@ -109,20 +107,20 @@ def test_gsm8k_correctness(config_filename):
 
         measured_metric = results["accuracy"]
         expected_metric = eval_config["accuracy_threshold"]
+        tol = eval_config.get("tolerance", 0.08)
 
         print(f"GSM8K Results for {eval_config['model_name']}:")
         print(f"  Measured metric: {measured_metric:.4f}")
         print(f"  Expected metric: {expected_metric:.4f}")
-        print(f"  Tolerance: {TOL:.4f}")
+        print(f"  Tolerance: {tol:.4f}")
         print(f"  Questions: {results['num_questions']}")
         print(f"  Invalid rate: {results['invalid_rate']:.3f}")
         print(f"  Latency: {results['latency']:.1f}s")
         print(f"  QPS: {results['questions_per_second']:.1f}")
 
-        # Verify metric is within tolerance
-        assert measured_metric >= expected_metric - TOL, (
+        assert measured_metric >= expected_metric - tol, (
             f"GSM8K metric too low: {measured_metric:.4f} < "
-            f"{expected_metric:.4f} - {TOL:.4f} = {expected_metric - TOL:.4f}"
+            f"{expected_metric:.4f} - {tol:.4f} = {expected_metric - tol:.4f}"
         )
 
         print(f"✅ GSM8K test passed for {eval_config['model_name']}")

vllm/config/vllm.py

@@ -701,6 +701,25 @@ class VllmConfig:
                 self.model_config, self.load_config
             )
 
+        if (
+            self.quant_config is not None
+            and self.model_config is not None
+            and hasattr(self.quant_config, "use_deep_gemm")
+            and self.quant_config.use_deep_gemm is None
+        ):
+            from vllm.utils.deep_gemm import should_auto_disable_deep_gemm
+
+            model_type = getattr(self.model_config.hf_text_config, "model_type", None)
+            if should_auto_disable_deep_gemm(model_type):
+                self.quant_config.use_deep_gemm = False
+                logger.warning_once(
+                    "Auto-disabled DeepGemm for model_type=%s on Blackwell. "
+                    "DeepGemm E8M0 scale format causes accuracy degradation "
+                    "for this architecture. Falling back to CUTLASS. "
+                    "To disable DeepGemm globally, set VLLM_USE_DEEP_GEMM=0.",
+                    model_type,
+                )
+
         from vllm.v1.executor.abstract import Executor
 
         executor_backend = self.parallel_config.distributed_executor_backend

vllm/model_executor/layers/quantization/fp8.py

@@ -129,6 +129,7 @@ class Fp8Config(QuantizationConfig):
                     f"{activation_scheme} activation scheme."
                 )
         self.weight_block_size = weight_block_size
+        self.use_deep_gemm: bool | None = None
 
     @classmethod
     def get_name(cls) -> QuantizationMethods:
@@ -276,7 +277,10 @@ class Fp8LinearMethod(LinearMethodBase):
         self.marlin_input_dtype = None
 
         self.use_aiter_and_is_supported = rocm_aiter_ops.is_linear_fp8_enabled()
-        self.use_deep_gemm = is_deep_gemm_supported()
+        if self.quant_config.use_deep_gemm is not None:
+            self.use_deep_gemm = self.quant_config.use_deep_gemm
+        else:
+            self.use_deep_gemm = is_deep_gemm_supported()
 
         self.weight_block_size = self.quant_config.weight_block_size
         self.block_quant = self.weight_block_size is not None
@@ -311,6 +315,7 @@ class Fp8LinearMethod(LinearMethodBase):
                 act_quant_group_shape=GroupShape(1, self.weight_block_size[0]),
                 cutlass_block_fp8_supported=self.cutlass_block_fp8_supported,
                 use_aiter_and_is_supported=self.use_aiter_and_is_supported,
+                use_deep_gemm=self.use_deep_gemm,
             )
 
     def create_weights(
@@ -428,7 +433,7 @@ class Fp8LinearMethod(LinearMethodBase):
         else:
             layer.input_scale = None
 
-        if self.block_quant:
+        if self.block_quant and self.use_deep_gemm:
             maybe_post_process_fp8_weight_block(layer)
 
     def apply(

vllm/model_executor/layers/quantization/input_quant_fp8.py

@@ -91,6 +91,7 @@ class QuantFP8(CustomOp):
 
         if (
             self.is_group_quant
+            and self.use_ue8m0
             and self.use_deep_gemm_supported
             and (DeepGemmQuantScaleFMT.from_oracle() == DeepGemmQuantScaleFMT.UE8M0)
         ):

vllm/model_executor/layers/quantization/utils/fp8_utils.py

@@ -359,10 +359,14 @@ class W8A8BlockFp8LinearOp:
         act_quant_group_shape: GroupShape,
         cutlass_block_fp8_supported: bool = CUTLASS_BLOCK_FP8_SUPPORTED,
         use_aiter_and_is_supported: bool = False,
+        use_deep_gemm: bool | None = None,
     ):
         self.weight_group_shape = weight_group_shape
         self.act_quant_group_shape = act_quant_group_shape
-        self.is_deep_gemm_supported = is_deep_gemm_supported()
+        if use_deep_gemm is not None:
+            self.is_deep_gemm_supported = use_deep_gemm
+        else:
+            self.is_deep_gemm_supported = is_deep_gemm_supported()
         self.is_hopper = current_platform.is_device_capability(90)
         self.use_deep_gemm_e8m0 = is_deep_gemm_e8m0_used()
         self.is_flashinfer_supported = is_flashinfer_fp8_blockscale_gemm_supported()

vllm/utils/deep_gemm.py

@@ -23,6 +23,24 @@ from vllm.platforms import current_platform
 from vllm.utils.import_utils import has_deep_gemm
 from vllm.utils.math_utils import cdiv
 
+_DEEPGEMM_BLACKWELL_EXCLUDED_MODEL_TYPES: set[str] = {
+    "qwen3_5_text",
+    "qwen3_5_moe_text",
+}
+
+
+def should_auto_disable_deep_gemm(model_type: str | None) -> bool:
+    """Check if DeepGemm should be auto-disabled for this model on Blackwell.
+
+    Returns True if the model is known to have accuracy degradation with
+    DeepGemm's E8M0 scale format on Blackwell GPUs (SM100+).
+    """
+    if model_type is None:
+        return False
+    if not current_platform.is_device_capability_family(100):
+        return False
+    return model_type in _DEEPGEMM_BLACKWELL_EXCLUDED_MODEL_TYPES
+
 
 class DeepGemmQuantScaleFMT(Enum):
     # Float32 scales in Float32 tensor