[TRTLLM-15433][chore] BREAKING: Remove all WIDEEP files

.pre-commit-config.yaml

@@ -264,13 +264,8 @@ common-files: &common_files |
         tensorrt_llm/_torch/modules/fused_moe/fused_moe_triton.py |
         tensorrt_llm/_torch/modules/fused_moe/fused_moe_trtllm_gen.py |
         tensorrt_llm/_torch/modules/fused_moe/fused_moe_vanilla.py |
-        tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py |
         tensorrt_llm/_torch/modules/fused_moe/interface.py |
         tensorrt_llm/_torch/modules/fused_moe/moe_load_balancer.py |
-        tensorrt_llm/_torch/modules/fused_moe/ops/__init__.py |
-        tensorrt_llm/_torch/modules/fused_moe/ops/moe_op.py |
-        tensorrt_llm/_torch/modules/fused_moe/ops/moe_op_cutlass.py |
-        tensorrt_llm/_torch/modules/fused_moe/ops/moe_op_deepgemm.py |
         tensorrt_llm/_torch/modules/fused_moe/quantization.py |
         tensorrt_llm/_torch/modules/fused_moe/routing.py |
         tensorrt_llm/_torch/modules/gated_mlp.py |
@@ -1038,13 +1033,8 @@ legacy-files: &legacy_files |
         tensorrt_llm/_torch/modules/fused_moe/fused_moe_triton.py |
         tensorrt_llm/_torch/modules/fused_moe/fused_moe_trtllm_gen.py |
         tensorrt_llm/_torch/modules/fused_moe/fused_moe_vanilla.py |
-        tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py |
         tensorrt_llm/_torch/modules/fused_moe/interface.py |
         tensorrt_llm/_torch/modules/fused_moe/moe_load_balancer.py |
-        tensorrt_llm/_torch/modules/fused_moe/ops/__init__.py |
-        tensorrt_llm/_torch/modules/fused_moe/ops/moe_op.py |
-        tensorrt_llm/_torch/modules/fused_moe/ops/moe_op_cutlass.py |
-        tensorrt_llm/_torch/modules/fused_moe/ops/moe_op_deepgemm.py |
         tensorrt_llm/_torch/modules/fused_moe/quantization.py |
         tensorrt_llm/_torch/modules/fused_moe/routing.py |
         tensorrt_llm/_torch/modules/gated_mlp.py |

examples/layer_wise_benchmarks/run.py

@@ -244,7 +244,7 @@ def comma_separated_floats(s):
     ctx_attn_workspace = torch.empty((0,), device="cuda", dtype=torch.int8)
     with mock.patch.dict(
         os.environ,
-        {"TRTLLM_FORCE_ALLTOALL_METHOD": "NotEnabled", "TRTLLM_FORCE_COMM_METHOD": "ALLGATHER"},
+        {"TRTLLM_FORCE_COMM_METHOD": "ALLGATHER"},
         clear=False,
     ):
         ctx_runner = Runner(

legacy-files.txt

@@ -256,13 +256,8 @@ tensorrt_llm/_torch/modules/fused_moe/fused_moe_deepgemm.py
 tensorrt_llm/_torch/modules/fused_moe/fused_moe_triton.py
 tensorrt_llm/_torch/modules/fused_moe/fused_moe_trtllm_gen.py
 tensorrt_llm/_torch/modules/fused_moe/fused_moe_vanilla.py
-tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py
 tensorrt_llm/_torch/modules/fused_moe/interface.py
 tensorrt_llm/_torch/modules/fused_moe/moe_load_balancer.py
-tensorrt_llm/_torch/modules/fused_moe/ops/__init__.py
-tensorrt_llm/_torch/modules/fused_moe/ops/moe_op.py
-tensorrt_llm/_torch/modules/fused_moe/ops/moe_op_cutlass.py
-tensorrt_llm/_torch/modules/fused_moe/ops/moe_op_deepgemm.py
 tensorrt_llm/_torch/modules/fused_moe/quantization.py
 tensorrt_llm/_torch/modules/fused_moe/routing.py
 tensorrt_llm/_torch/modules/gated_mlp.py

pyproject.toml

@@ -313,13 +313,8 @@ exclude = [
     "tensorrt_llm/_torch/modules/fused_moe/fused_moe_triton.py",
     "tensorrt_llm/_torch/modules/fused_moe/fused_moe_trtllm_gen.py",
     "tensorrt_llm/_torch/modules/fused_moe/fused_moe_vanilla.py",
-    "tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py",
     "tensorrt_llm/_torch/modules/fused_moe/interface.py",
     "tensorrt_llm/_torch/modules/fused_moe/moe_load_balancer.py",
-    "tensorrt_llm/_torch/modules/fused_moe/ops/__init__.py",
-    "tensorrt_llm/_torch/modules/fused_moe/ops/moe_op.py",
-    "tensorrt_llm/_torch/modules/fused_moe/ops/moe_op_cutlass.py",
-    "tensorrt_llm/_torch/modules/fused_moe/ops/moe_op_deepgemm.py",
     "tensorrt_llm/_torch/modules/fused_moe/quantization.py",
     "tensorrt_llm/_torch/modules/fused_moe/routing.py",
     "tensorrt_llm/_torch/modules/gated_mlp.py",

ruff-legacy.toml

@@ -273,13 +273,8 @@ include = [
     "tensorrt_llm/_torch/modules/fused_moe/fused_moe_triton.py",
     "tensorrt_llm/_torch/modules/fused_moe/fused_moe_trtllm_gen.py",
     "tensorrt_llm/_torch/modules/fused_moe/fused_moe_vanilla.py",
-    "tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py",
     "tensorrt_llm/_torch/modules/fused_moe/interface.py",
     "tensorrt_llm/_torch/modules/fused_moe/moe_load_balancer.py",
-    "tensorrt_llm/_torch/modules/fused_moe/ops/__init__.py",
-    "tensorrt_llm/_torch/modules/fused_moe/ops/moe_op.py",
-    "tensorrt_llm/_torch/modules/fused_moe/ops/moe_op_cutlass.py",
-    "tensorrt_llm/_torch/modules/fused_moe/ops/moe_op_deepgemm.py",
     "tensorrt_llm/_torch/modules/fused_moe/quantization.py",
     "tensorrt_llm/_torch/modules/fused_moe/routing.py",
     "tensorrt_llm/_torch/modules/gated_mlp.py",

tensorrt_llm/_torch/models/modeling_deepseekv3.py

@@ -63,7 +63,6 @@
 from ..modules.fused_moe import (DeepSeekV3MoeRoutingMethod,
                                  MoEWeightLoadingMode, create_moe,
                                  is_moe_weight_owner)
-from ..modules.fused_moe.fused_moe_wide_ep import WideEPMoE
 from ..modules.mla import MLA
 
 # isort: off
@@ -1151,9 +1150,6 @@ def compute_routed_output(self, hidden_states, hidden_states_fp4,
             output_dtype=hidden_states.dtype,
             all_rank_num_tokens=all_rank_num_tokens,
             use_dp_padding=use_dp_padding,
-            **({
-                "alltoall_result_do_sum": False
-            } if isinstance(self.experts, WideEPMoE) else {}),
         )
 
         return routed_output

tensorrt_llm/_torch/models/modeling_deepseekv4.py

@@ -76,7 +76,6 @@
     resolve_moe_cls,
 )
 from ..modules.fused_moe.fused_moe_deepgemm import DeepGemmFusedMoE
-from ..modules.fused_moe.fused_moe_wide_ep import WideEPMoE
 from ..modules.gated_mlp import GatedMLP
 from ..modules.linear import Linear, TensorParallelMode, WeightsLoadingConfig
 from ..modules.mhc.hyper_connection import HCHead, HCState, mHC
@@ -1697,7 +1696,6 @@ def compute_routed_output(
             output_dtype=hidden_states.dtype,
             all_rank_num_tokens=all_rank_num_tokens,
             use_dp_padding=use_dp_padding,
-            **({"alltoall_result_do_sum": False} if isinstance(self.experts, WideEPMoE) else {}),
         )
 
         return routed_output

tensorrt_llm/_torch/modules/fused_moe/MOE_DEVELOPER_GUIDE.md

@@ -112,15 +112,16 @@ The codebase is transitioning between two architectures:
 | Entry | `XXFusedMoE` (e.g., `CutlassFusedMoE`) | `ConfigurableMoE` + `XXBackend` + `MoEScheduler` |
 | Communication | Embedded inside each backend | Separated into `communication/` (or fused into kernel for `FUSED_COMM`) |
 | Forward execution | Inline in backend | `MoEScheduler` (`moe_scheduler.py`) |
-| EPLB | Only in WideEPMoE | Available to all backends |
+| EPLB | Not supported | Available on EPLB-capable backends |
 | Status | Being replaced | Active development |
 
 ConfigurableMoE currently supports these backends (`create_moe.py`):
-- `CutlassFusedMoE`, `TRTLLMGenFusedMoE`, `DeepGemmFusedMoE`, `CuteDslFusedMoE`, `DenseGEMMFusedMoE`, `MegaMoEDeepGemm`, `MegaMoECuteDsl`
+- `CutlassFusedMoE`, `TRTLLMGenFusedMoE`, `DeepGemmFusedMoE`, `CuteDslFusedMoE`,
+  `CuteDslB12xFusedMoE`, `DenseGEMMFusedMoE`, `MegaMoEDeepGemm`,
+  `MegaMoECuteDsl`, `MarlinFusedMoE`
 
 Still on old path (standalone, with embedded communication):
 - `TritonFusedMoE`, `VanillaMoE`
-- `WideEPMoE` — deprecated, the `WIDEEP` backend can no longer be selected
 
 **Rule: All new features should target ConfigurableMoE + Backend + Scheduler architecture.**
 
@@ -137,7 +138,7 @@ Still on old path (standalone, with embedded communication):
 | `impl_contract.py` | Selection vocabulary — `MoEProblem`, `MoEDeployment`, `MoEEnvironment`, `MoEEligibility`, `MoERejectReason`, `MoEResolutionReport` |
 | `impl_environment.py` | The only place that probes the machine (SM, optional wheels, env flags) and freezes the result |
 | `impl_identity.py` | `MoEImplId` / `MoEImplDescriptor` / registry — the stable one-id-per-leaf-class mechanism used after an implementation migrates |
-| `interface.py` | Base class `MoE` and enums (`MoEWeightLoadingMode`, `MoESchedulerKind`, `AlltoallMethodType`) |
+| `interface.py` | Base class `MoE` and enums (`MoEWeightLoadingMode`, `MoESchedulerKind`) |
 | `quantization.py` | Quantization method implementations (`FusedMoEMethod` subclasses: weight creation, loading, quant/dequant ops per quant mode) |
 | `routing.py` | Routing methods (`TopKRouting`, etc.) |
 | `moe_load_balancer.py` | EPLB implementation |
@@ -157,7 +158,6 @@ Still on old path (standalone, with embedded communication):
 | `mega_moe/mega_moe_cute_dsl.py` | `MegaMoECuteDsl` | SM100/SM103 | NVFP4 via ported CuteDSL `Sm100MegaMoEKernel` fused dispatch+FC1+act+FC2+combine kernel; requires CUDA 13 Cutlass DSL runtime (PR #14354) and NVSHMEM provider (hard gate); threads per-expert `fc31_alpha`/`fc2_alpha`/`fc1_norm_const` through the kernel ABI and supports SwiGLU clamp via `swiglu_limit`; default deepgemm graph (topk score folded before fc1-out quant, host `combine_output.sum(dim=1)`) | `FUSED_COMM` |
 | `fused_moe_marlin.py` | `MarlinFusedMoE` | SM89-SM99 | W4A16 NVFP4 on Ada/Hopper (BF16 activations + FP4 weights, fused single-launch `marlin_nvfp4_moe_gemm` kernel); supports attention-DP + EP via external comm (scheduler precomputes routing; dispatch payload is plain BF16, no activation scales); non-NVFP4 layers (e.g. unquantized MTP draft layers) degrade to Cutlass in `resolve_moe_impl`, recorded in the layer's `MoEResolutionReport`; no dynamic EPLB | `EXTERNAL_COMM` |
 | `fused_moe_triton.py` | `TritonFusedMoE` | SM90 only | GPT-OSS on Hopper (requires `swiglu_gptoss_style=True`) | (legacy path) |
-| `fused_moe_wide_ep.py` | `WideEPMoE` | All GPUs | Deprecated — `moe_resolution.py` raises on the `WIDEEP` backend literal. Wide EP and EPLB are available on the other backends: use `DEEPGEMM` for FP8 block-scale checkpoints, or `TRTLLM` / `CUTEDSL` / `CUTLASS` otherwise. Class kept for reference only | (legacy path) |
 | `fused_moe_vanilla.py` | `VanillaMoE` | All devices | Reference / debugging only | (legacy path) |
 
 ### Communication (`fused_moe/communication/`)
@@ -310,26 +310,21 @@ unconditionally, not here.
 
 Each backend's `can_implement(p, d)` classmethod declares what it supports. Source of truth: the `can_implement` classmethod in each backend file.
 
-| Quantization | Cutlass | TRTLLMGen | DeepGemm | DenseGEMM | CuteDSL | MegaMoE-DG | MegaMoE-CuteDSL | Triton | Marlin | WideEP (retired)† | Vanilla |
-|---|---|---|---|---|---|---|---|---|---|---|---|
-| Unquantized (BF16/FP16) | Y (SM80+) | Y (SM100/103, BF16, needs FlashInfer `trtllm_bf16_moe`)§ | N | N | N | N | N | Y (SM90, BF16) | N | Y | Y |
-| FP8 QDQ | Y (SM89+) | N | N | N | N | N | N | Y (SM90) | N | Y | Y |
-| FP8 Block Scales | Y (SM90, SM120) | Y (SM100/103) | Y (SM100/103) | N | N‡ | N | N | N | N | Y | Y |
-| NVFP4 | Y (SM100/103/120/121) | Y (SM100/103) | N | Y (SM100/103) | Y (SM100/103/120/121) | N | Y (SM100/103, cu13 cutlass-dsl + NVSHMEM provider; per-expert alpha/norm_const + SwiGLU clamp) | N | Y (SM89-SM99) | Y | Y |
-| W4A16 NVFP4 | Y (SM80+, dequant-on-the-fly) | N | N | N | Y (SM120/121 via `CuteDslB12xFusedMoE`, needs flashinfer) | N | N | N | Y (SM89-SM99, BF16) | N | Y |
-| W4A8 NVFP4 FP8 | N | Y (SM100/103) | N | N | N | N | N | N | N | N | N |
-| W4A16 MXFP4 | Y (SM90) | Y (SM100/103) | N | N | N | N | N | Y (SM90) | N | N | N |
-| W4A8 MXFP4 FP8 | Y (SM100/103) | Y (SM100/103) | N | N | N | N | N | Y (SM90) | N | N | N |
-| W4A8 MXFP4 MXFP8 | Y (SM100/103) | Y (SM100/103) | N | N | N | Y (SM100/103, requires `hidden_size % 512 == 0`) | N | N | N | N | N |
-| W8A8 MXFP8 MXFP8 | Y (SM100/103) | N | N | N | N | N | N | N | N | N | N |
-| W4A8 AWQ | Y (SM89/90) | N | N | N | N | N | N | N | N | N | N |
-| W8A16 | Y (SM80+) | N | N | N | N | N | N | N | N | N | N |
-| INT4 WoQ (W4AFP8) | N | N | N | N | N | N | N | N | N | Y | N |
-
-† The `WideEP` column is history, not an option. `resolve_moe_impl` raises on the
-`WIDEEP` literal, so nothing in that column can be selected — read a `Y` there as
-"the retired class implemented this", never as "you may request this". For
-`INT4 WoQ (W4AFP8)` that leaves the row without a selectable backend.
+| Quantization | Cutlass | TRTLLMGen | DeepGemm | DenseGEMM | CuteDSL | MegaMoE-DG | MegaMoE-CuteDSL | Triton | Marlin | Vanilla |
+|---|---|---|---|---|---|---|---|---|---|---|
+| Unquantized (BF16/FP16) | Y (SM80+) | Y (SM100/103, BF16, needs FlashInfer `trtllm_bf16_moe`)§ | N | N | N | N | N | Y (SM90, BF16) | N | Y |
+| FP8 QDQ | Y (SM89+) | N | N | N | N | N | N | Y (SM90) | N | Y |
+| FP8 Block Scales | Y (SM90, SM120) | Y (SM100/103) | Y (SM100/103) | N | N‡ | N | N | N | N | Y |
+| NVFP4 | Y (SM100/103/120/121) | Y (SM100/103) | N | Y (SM100/103) | Y (SM100/103/120/121) | N | Y (SM100/103, cu13 cutlass-dsl + NVSHMEM provider; per-expert alpha/norm_const + SwiGLU clamp) | N | Y (SM89-SM99) | Y |
+| W4A16 NVFP4 | Y (SM80+, dequant-on-the-fly) | N | N | N | Y (SM120/121 via `CuteDslB12xFusedMoE`, needs flashinfer) | N | N | N | Y (SM89-SM99, BF16) | Y |
+| W4A8 NVFP4 FP8 | N | Y (SM100/103) | N | N | N | N | N | N | N | N |
+| W4A16 MXFP4 | Y (SM90) | Y (SM100/103) | N | N | N | N | N | Y (SM90) | N | N |
+| W4A8 MXFP4 FP8 | Y (SM100/103) | Y (SM100/103) | N | N | N | N | N | Y (SM90) | N | N |
+| W4A8 MXFP4 MXFP8 | Y (SM100/103) | Y (SM100/103) | N | N | N | Y (SM100/103, requires `hidden_size % 512 == 0`) | N | N | N | N |
+| W8A8 MXFP8 MXFP8 | Y (SM100/103) | N | N | N | N | N | N | N | N | N |
+| W4A8 AWQ | Y (SM89/90) | N | N | N | N | N | N | N | N | N |
+| W8A16 | Y (SM80+) | N | N | N | N | N | N | N | N | N |
+| INT4 WoQ (W4AFP8) | N | N | N | N | N | N | N | N | N | N |
 
 § The unquantized `TRTLLMGenFusedMoE` path is not a TRTLLM-Gen kernel at all: it
 calls FlashInfer's `trtllm_bf16_moe` / `trtllm_bf16_routed_moe`, which is why it

tensorrt_llm/_torch/modules/fused_moe/__init__.py

@@ -7,7 +7,6 @@
 from .fused_moe_triton import TritonFusedMoE
 from .fused_moe_trtllm_gen import TRTLLMGenFusedMoE
 from .fused_moe_vanilla import VanillaMoE
-from .fused_moe_wide_ep import WideEPMoE
 from .impl_base import MoEImplBase
 from .interface import MoE, MoEWeightLoadingMode
 from .moe_load_balancer import (MoeLoadBalancer,
@@ -61,5 +60,4 @@
     "TritonFusedMoE",
     "TRTLLMGenFusedMoE",
     "VanillaMoE",
-    "WideEPMoE",
 ]

tensorrt_llm/_torch/modules/fused_moe/create_moe.py

@@ -18,7 +18,6 @@
 from .fused_moe_triton import TritonFusedMoE
 from .fused_moe_trtllm_gen import TRTLLMGenFusedMoE
 from .fused_moe_vanilla import VanillaMoE
-from .fused_moe_wide_ep import WideEPMoE
 from .interface import MoE, MoEWeightLoadingMode
 from .mega_moe import MegaMoECuteDsl, MegaMoEDeepGemm
 from .moe_load_balancer import get_moe_load_balancer
@@ -98,9 +97,6 @@ def create_moe_backend(
     Returns:
         MoE: MoE backend instance
     """
-    if moe_cls is WideEPMoE:
-        raise ValueError(WIDEEP_DEPRECATION_MESSAGE)
-
     shapes = derive_moe_layer_shapes(
         model_config,
         num_experts=num_experts,
@@ -499,7 +495,7 @@ def create_moe(
             communication_method=communication_method,
         )
 
-    # WideEPMoE, TritonFusedMoE and VanillaMoE are not wrapped by ConfigurableMoE
+    # TritonFusedMoE and VanillaMoE are not wrapped by ConfigurableMoE
     # and own their communication and forward paths.
     if communication_method is not None:
         raise ValueError("communication_method requires ConfigurableMoE.")