diff --git a/docs/design/moe_kernel_features.md b/docs/design/moe_kernel_features.md index e2fd5645f6..231bca3646 100644 --- a/docs/design/moe_kernel_features.md +++ b/docs/design/moe_kernel_features.md @@ -86,7 +86,7 @@ To be used with a particular `FusedMoEPrepareAndFinalizeModular` subclass, MoE k | cutlass_fp4 | standard,
batched | nvfp4 | A,T | silu | Y | Y | [`CutlassExpertsFp4`][vllm.model_executor.layers.fused_moe.cutlass_moe.CutlassExpertsFp4] | | cutlass_fp8 | standard,
batched | fp8 | A,T | silu, gelu | Y | Y | [`CutlassExpertsFp8`][vllm.model_executor.layers.fused_moe.cutlass_moe.CutlassExpertsFp8],
[`CutlasBatchedExpertsFp8`][vllm.model_executor.layers.fused_moe.cutlass_moe.CutlassBatchedExpertsFp8] | | flashinfer | standard | nvfp4,
fp8 | T | 5 | N | Y | [`FlashInferExperts`][vllm.model_executor.layers.fused_moe.flashinfer_cutlass_moe.FlashInferExperts] | -| gpt oss triton | standard | N/A | N/A | 5 | Y | Y | [`triton_kernel_fused_experts`][vllm.model_executor.layers.fused_moe.gpt_oss_triton_kernels_moe.triton_kernel_fused_experts],
[`OAITritonExperts`][vllm.model_executor.layers.fused_moe.gpt_oss_triton_kernels_moe.OAITritonExperts] | +| gpt oss triton | standard | N/A | N/A | 5 | Y | Y | [`triton_kernel_fused_experts`][vllm.model_executor.layers.fused_moe.experts.gpt_oss_triton_kernels_moe.triton_kernel_fused_experts],
[`OAITritonExperts`][vllm.model_executor.layers.fused_moe.experts.gpt_oss_triton_kernels_moe.OAITritonExperts] | | marlin | standard,
batched | 3 / N/A | 3 / N/A | silu,
swigluoai | Y | Y | [`fused_marlin_moe`][vllm.model_executor.layers.fused_moe.fused_marlin_moe.fused_marlin_moe],
[`MarlinExperts`][vllm.model_executor.layers.fused_moe.fused_marlin_moe.MarlinExperts],
[`BatchedMarlinExperts`][vllm.model_executor.layers.fused_moe.fused_marlin_moe.BatchedMarlinExperts] | | trtllm | standard | mxfp4,
nvfp4 | G(16),G(32) | 5 | N | Y | [`TrtLlmMxfp4ExpertsMonolithic`][vllm.model_executor.layers.fused_moe.experts.trtllm_mxfp4_moe.TrtLlmMxfp4ExpertsMonolithic],
[`TrtLlmMxfp4ExpertsModular`][vllm.model_executor.layers.fused_moe.experts.trtllm_mxfp4_moe.TrtLlmMxfp4ExpertsModular],
[`TrtLlmNvFp4ExpertsMonolithic`][vllm.model_executor.layers.fused_moe.experts.trtllm_nvfp4_moe.TrtLlmNvFp4ExpertsMonolithic],
[`TrtLlmNvfp4ExpertsModular`][vllm.model_executor.layers.fused_moe.experts.trtllm_nvfp4_moe.TrtLlmNvFp4ExpertsModular] | | rocm aiter moe | standard | mxfp4,
fp8 | G(32),G(128),A,T | silu, gelu,
swigluoai | Y | N | `rocm_aiter_fused_experts`,
`AiterExperts` | diff --git a/tests/kernels/moe/test_gpt_oss_triton_kernels.py b/tests/kernels/moe/test_gpt_oss_triton_kernels.py index 032b4fc047..c61004acaa 100644 --- a/tests/kernels/moe/test_gpt_oss_triton_kernels.py +++ b/tests/kernels/moe/test_gpt_oss_triton_kernels.py @@ -25,7 +25,7 @@ from triton_kernels.tensor_details import layout from triton_kernels.testing import assert_close from vllm.model_executor.layers.fused_moe.config import mxfp4_w4a16_moe_quant_config -from vllm.model_executor.layers.fused_moe.gpt_oss_triton_kernels_moe import ( +from vllm.model_executor.layers.fused_moe.experts.gpt_oss_triton_kernels_moe import ( triton_kernel_moe_forward, ) from vllm.utils.math_utils import round_up diff --git a/tests/kernels/moe/test_modular_oai_triton_moe.py b/tests/kernels/moe/test_modular_oai_triton_moe.py index b071e72daf..589d90d1ec 100644 --- a/tests/kernels/moe/test_modular_oai_triton_moe.py +++ b/tests/kernels/moe/test_modular_oai_triton_moe.py @@ -29,7 +29,7 @@ from vllm.model_executor.layers.fused_moe.all2all_utils import ( maybe_make_prepare_finalize, ) from vllm.model_executor.layers.fused_moe.config import mxfp4_w4a16_moe_quant_config -from vllm.model_executor.layers.fused_moe.gpt_oss_triton_kernels_moe import ( +from vllm.model_executor.layers.fused_moe.experts.gpt_oss_triton_kernels_moe import ( OAITritonExperts, UnfusedOAITritonExperts, ) diff --git a/tests/kernels/quantization/test_mxfp4_triton_ep.py b/tests/kernels/quantization/test_mxfp4_triton_ep.py index 045bc63de9..db00d74bd6 100644 --- a/tests/kernels/quantization/test_mxfp4_triton_ep.py +++ b/tests/kernels/quantization/test_mxfp4_triton_ep.py @@ -29,18 +29,22 @@ class TestTritonMoeForwardExpertMap: torch.tensor([0, -1, 1, -1], device=device) if expert_map_present else None ) + from vllm.utils.import_utils import import_triton_kernels + + import_triton_kernels() + with ( patch("triton_kernels.topk.topk") as mock_topk, patch( - "vllm.model_executor.layers.fused_moe." + "vllm.model_executor.layers.fused_moe.experts." "gpt_oss_triton_kernels_moe.make_routing_data" ) as mock_make_routing, patch( - "vllm.model_executor.layers.fused_moe." + "vllm.model_executor.layers.fused_moe.experts." "gpt_oss_triton_kernels_moe.triton_kernel_fused_experts" ) as mock_fused_experts, ): - from vllm.model_executor.layers.fused_moe.gpt_oss_triton_kernels_moe import ( # noqa: E501 + from vllm.model_executor.layers.fused_moe.experts.gpt_oss_triton_kernels_moe import ( # noqa: E501 triton_kernel_moe_forward, ) diff --git a/vllm/lora/layers/fused_moe.py b/vllm/lora/layers/fused_moe.py index 01efe3e473..835bffe58c 100644 --- a/vllm/lora/layers/fused_moe.py +++ b/vllm/lora/layers/fused_moe.py @@ -19,6 +19,9 @@ from vllm.model_executor.layers.fused_moe import FusedMoE from vllm.model_executor.layers.fused_moe.config import ( _get_config_dtype_str, ) +from vllm.model_executor.layers.fused_moe.experts.gpt_oss_triton_kernels_moe import ( + UnfusedOAITritonExperts, +) from vllm.model_executor.layers.fused_moe.fused_marlin_moe import ( MarlinExperts, ) @@ -28,9 +31,6 @@ from vllm.model_executor.layers.fused_moe.fused_moe import ( from vllm.model_executor.layers.fused_moe.fused_moe_modular_method import ( FusedMoEModularMethod, ) -from vllm.model_executor.layers.fused_moe.gpt_oss_triton_kernels_moe import ( - UnfusedOAITritonExperts, -) from vllm.model_executor.layers.fused_moe.modular_kernel import ( FusedMoEKernel, ) diff --git a/vllm/model_executor/layers/fused_moe/gpt_oss_triton_kernels_moe.py b/vllm/model_executor/layers/fused_moe/experts/gpt_oss_triton_kernels_moe.py similarity index 100% rename from vllm/model_executor/layers/fused_moe/gpt_oss_triton_kernels_moe.py rename to vllm/model_executor/layers/fused_moe/experts/gpt_oss_triton_kernels_moe.py diff --git a/vllm/model_executor/layers/fused_moe/oracle/mxfp4.py b/vllm/model_executor/layers/fused_moe/oracle/mxfp4.py index e7cfb881c1..917d474fc9 100644 --- a/vllm/model_executor/layers/fused_moe/oracle/mxfp4.py +++ b/vllm/model_executor/layers/fused_moe/oracle/mxfp4.py @@ -101,7 +101,7 @@ def backend_to_kernel_cls( return [FlashInferExperts] elif backend == Mxfp4MoeBackend.TRITON: - from vllm.model_executor.layers.fused_moe.gpt_oss_triton_kernels_moe import ( + from vllm.model_executor.layers.fused_moe.experts.gpt_oss_triton_kernels_moe import ( # noqa: E501 OAITritonExperts, OAITritonMxfp4ExpertsMonolithic, ) @@ -110,7 +110,7 @@ def backend_to_kernel_cls( return [OAITritonMxfp4ExpertsMonolithic, OAITritonExperts] elif backend == Mxfp4MoeBackend.TRITON_UNFUSED: - from vllm.model_executor.layers.fused_moe.gpt_oss_triton_kernels_moe import ( + from vllm.model_executor.layers.fused_moe.experts.gpt_oss_triton_kernels_moe import ( # noqa: E501 UnfusedOAITritonExperts, ) diff --git a/vllm/model_executor/layers/quantization/quark/quark_moe.py b/vllm/model_executor/layers/quantization/quark/quark_moe.py index 4ad421aa80..2bab66709d 100644 --- a/vllm/model_executor/layers/quantization/quark/quark_moe.py +++ b/vllm/model_executor/layers/quantization/quark/quark_moe.py @@ -1591,7 +1591,7 @@ class QuarkOCP_MX_MoEMethod_OSS(QuarkOCP_MX_MoEMethod): "EPLB not supported for `QuarkW4MXFp4MoEMethod_OSS` yet." ) - from vllm.model_executor.layers.fused_moe.gpt_oss_triton_kernels_moe import ( # noqa: E501 + from vllm.model_executor.layers.fused_moe.experts.gpt_oss_triton_kernels_moe import ( # noqa: E501 triton_kernel_moe_forward, )