[Quantization] Auto-disable DeepGEMM for Kolibri 1 on SM12x (#61274)

Signed-off-by: Luca Motz <luca.motz@icloud.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

vllm/config/vllm.py

@@ -1547,8 +1547,6 @@ class VllmConfig:
                 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,
                 )

vllm/model_executor/layers/fused_moe/oracle/fp8.py

@@ -9,6 +9,7 @@ import torch
 import vllm.model_executor.layers.fused_moe.modular_kernel as mk
 from vllm import envs
 from vllm._aiter_ops import rocm_aiter_ops
+from vllm.config import get_current_vllm_config
 from vllm.config.kernel import MoEBackend
 from vllm.logger import init_logger
 from vllm.model_executor.layers.fused_moe.all2all_utils import (
@@ -40,6 +41,7 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import (
 )
 from vllm.model_executor.utils import replace_parameter
 from vllm.platforms import current_platform
+from vllm.utils.deep_gemm import should_auto_disable_deep_gemm
 from vllm.utils.math_utils import round_up
 
 logger = init_logger(__name__)
@@ -492,6 +494,13 @@ def select_fp8_moe_backend(
             return _return_or_raise(
                 backend, config, weight_key, activation_key, activation_format
             )
+    elif (model_config := get_current_vllm_config().model_config) is not None and (
+        should_auto_disable_deep_gemm(
+            getattr(model_config.hf_text_config, "model_type", None)
+        )
+    ):
+        AVAILABLE_BACKENDS.remove(Fp8MoeBackend.DEEPGEMM)
+        AVAILABLE_BACKENDS.remove(Fp8MoeBackend.BATCHED_DEEPGEMM)
 
     # Handle explicit AITER FP8 configuration.
     if envs.is_set("VLLM_ROCM_USE_AITER") or envs.is_set("VLLM_ROCM_USE_AITER_MOE"):

vllm/utils/deep_gemm.py

@@ -24,9 +24,13 @@ 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",
+# Model type -> Blackwell device capability families to auto-disable DeepGemm on.
+_DEEPGEMM_BLACKWELL_EXCLUDED_MODEL_TYPES: dict[str, tuple[int, ...]] = {
+    # Accuracy
+    "qwen3_5_text": (100, 120),
+    "qwen3_5_moe_text": (100, 120),
+    # Latency
+    "kolibri1": (120,),
 }
 
 
@@ -39,17 +43,15 @@ def get_paged_mqa_page_sizes() -> tuple[int, ...]:
 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+).
+    Returns True if DeepGemm is known to hurt this model's accuracy or
+    latency on the current Blackwell GPU family.
     """
     if model_type is None:
         return False
-    if not (
-        current_platform.is_device_capability_family(100)
-        or current_platform.is_device_capability_family(120)
-    ):
-        return False
-    return model_type in _DEEPGEMM_BLACKWELL_EXCLUDED_MODEL_TYPES
+    return any(
+        current_platform.is_device_capability_family(family)
+        for family in _DEEPGEMM_BLACKWELL_EXCLUDED_MODEL_TYPES.get(model_type, ())
+    )
 
 
 class DeepGemmQuantScaleFMT(Enum):