diff --git a/.buildkite/test_areas/lm_eval.yaml b/.buildkite/test_areas/lm_eval.yaml index 217fc5665c..a5beb5ea36 100644 --- a/.buildkite/test_areas/lm_eval.yaml +++ b/.buildkite/test_areas/lm_eval.yaml @@ -109,6 +109,7 @@ steps: - image-build-amd commands: - export VLLM_WORKER_MULTIPROC_METHOD=spawn + - export PYTORCH_ROCM_ARCH=gfx942 # Limit Quark compilation to save time - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-mi3xx.txt - label: MoE Refactor Integration Test (H100 - TEMPORARY) diff --git a/vllm/model_executor/layers/fused_moe/experts/nvfp4_emulation_moe.py b/vllm/model_executor/layers/fused_moe/experts/nvfp4_emulation_moe.py index de5b45ccb8..d7ed53612e 100644 --- a/vllm/model_executor/layers/fused_moe/experts/nvfp4_emulation_moe.py +++ b/vllm/model_executor/layers/fused_moe/experts/nvfp4_emulation_moe.py @@ -21,7 +21,6 @@ from vllm.model_executor.layers.fused_moe.config import ( FusedMoEQuantConfig, ) from vllm.model_executor.layers.fused_moe.experts.triton_moe import TritonExperts -from vllm.model_executor.layers.fused_moe.utils import moe_kernel_quantize_input from vllm.model_executor.layers.quantization.utils.nvfp4_emulation_utils import ( dequantize_to_dtype, ) @@ -135,14 +134,6 @@ class Nvfp4QuantizationEmulationTritonExperts(TritonExperts): swizzle=False, ) - hidden_states, _ = moe_kernel_quantize_input( - A=hidden_states, - A_scale=self.quant_config.a1_gscale, - quant_dtype="nvfp4", - per_act_token_quant=False, - quantization_emulation=True, - ) - # Activation quantization/dequantization is deferred to # `moe_kernel_quantize_input` in TritonExperts.apply. super().apply( diff --git a/vllm/model_executor/layers/fused_moe/experts/ocp_mx_emulation_moe.py b/vllm/model_executor/layers/fused_moe/experts/ocp_mx_emulation_moe.py index feb8c2ea76..b29e2fde01 100644 --- a/vllm/model_executor/layers/fused_moe/experts/ocp_mx_emulation_moe.py +++ b/vllm/model_executor/layers/fused_moe/experts/ocp_mx_emulation_moe.py @@ -21,7 +21,6 @@ from vllm.model_executor.layers.fused_moe.config import ( FusedMoEQuantConfig, ) from vllm.model_executor.layers.fused_moe.experts.triton_moe import TritonExperts -from vllm.model_executor.layers.fused_moe.utils import moe_kernel_quantize_input from vllm.model_executor.layers.quantization.utils.mxfp4_utils import dequant_mxfp4 from vllm.model_executor.layers.quantization.utils.mxfp6_utils import dequant_mxfp6 from vllm.model_executor.layers.quantization.utils.ocp_mx_utils import ( @@ -155,16 +154,6 @@ class OCP_MXQuantizationEmulationTritonExperts(TritonExperts): w2, self.w2_scale_val, hidden_states.dtype ) - # Apply activation QDQ if needed by the OCP MX scheme - hidden_states, _ = moe_kernel_quantize_input( - A=hidden_states, - A_scale=None, - quant_dtype=self.quant_config.quant_dtype, - per_act_token_quant=False, - ocp_mx_scheme=self.ocp_mx_scheme, - quantization_emulation=True, - ) - # Activation quantization/dequantization is deferred to # `moe_kernel_quantize_input` in TritonExperts.apply. super().apply( diff --git a/vllm/model_executor/layers/fused_moe/experts/triton_moe.py b/vllm/model_executor/layers/fused_moe/experts/triton_moe.py index 0d9b43658f..abe31e017d 100644 --- a/vllm/model_executor/layers/fused_moe/experts/triton_moe.py +++ b/vllm/model_executor/layers/fused_moe/experts/triton_moe.py @@ -245,7 +245,7 @@ class TritonExperts(LoRAExpertsMixin, mk.FusedMoEExpertsModular): lora_unquantized_hidden_states = hidden_states hidden_states, a1q_scale = moe_kernel_quantize_input( hidden_states, - self.a1_scale, + self.a1_scale or self.a1_gscale, self.quant_dtype, self.per_act_token_quant, self.block_shape, diff --git a/vllm/model_executor/layers/fused_moe/utils.py b/vllm/model_executor/layers/fused_moe/utils.py index b8c84ad2af..8866b4f09f 100644 --- a/vllm/model_executor/layers/fused_moe/utils.py +++ b/vllm/model_executor/layers/fused_moe/utils.py @@ -296,6 +296,7 @@ def moe_kernel_quantize_input( if not quantization_emulation: return _nvfp4_quantize(A, A_scale, is_sf_swizzled_layout=is_scale_swizzled) else: + assert A_scale is not None A = ref_nvfp4_quant_dequant(A, A_scale, block_size=16) return A, None elif quant_dtype == "mxfp4":