[None][refactor] overhaul NVLink one-sided all-to-all

.claude/skills/trtllm-moe-develop/SKILL.md

@@ -673,7 +673,7 @@ Prefer the unified MoE tests:
 - Communication changes: `pytest tests/unittest/_torch/moe/test_moe_comm.py -k '<strategy>'`.
 - Routing changes: `pytest tests/unittest/_torch/moe/test_moe_routing.py -k '<routing>'`.
 - Load balancer changes: `pytest tests/unittest/_torch/moe/test_moe_load_balancer.py -k '<case>'`.
-- Multi-GPU EP/all-to-all behavior: `pytest tests/unittest/_torch/moe/multi_gpu/test_moe_a2a.py -k '<case>'`.
+- Multi-GPU EP/all-to-all behavior: `pytest tests/unittest/_torch/moe/multi_gpu/test_nvlink_one_sided.py -k '<case>'`.
 
 When GPU resources are required, use the TRT-LLM GPU allocation/test-runner
 skills first and record skipped tests with reasons.

.claude/skills/trtllm-moe-develop/references/moe-canonical-code-examples.md

@@ -371,7 +371,7 @@ Use these examples when wrapper forward policy grows complicated:
     should move into the scheduler.
 - `tests/unittest/_torch/moe/test_moe_module.py`
   - Module-level multi-GPU, chunking, routing, and EPLB cases.
-- `tests/unittest/_torch/moe/multi_gpu/test_moe_a2a.py`
+- `tests/unittest/_torch/moe/multi_gpu/test_nvlink_one_sided.py`
   - Multi-GPU all-to-all behavior when relevant.
 
 Good uses:

.pre-commit-config.yaml

@@ -273,7 +273,6 @@ common-files: &common_files |
         tensorrt_llm/_torch/modules/triton_linear.py |
         tensorrt_llm/_torch/moe/expert_statistic.py |
         tensorrt_llm/_torch/moe/fused_moe/__init__.py |
-        tensorrt_llm/_torch/moe/fused_moe/communication/moe_alltoall.py |
         tensorrt_llm/_torch/moe/fused_moe/create_moe.py |
         tensorrt_llm/_torch/moe/fused_moe/deep_ep_utils.py |
         tensorrt_llm/_torch/moe/fused_moe/fused_moe_cute_dsl.py |
@@ -582,7 +581,6 @@ common-files: &common_files |
         tests/unittest/_torch/modeling/test_modeling_vila.py |
         tests/unittest/_torch/modules/test_group_rmn_norm.py |
         tests/unittest/_torch/modules/test_triton_linear.py |
-        tests/unittest/_torch/moe/multi_gpu/test_moe_a2a.py |
         tests/unittest/_torch/moe/test_fused_moe.py |
         tests/unittest/_torch/moe/test_moe_host_sharer.py |
         tests/unittest/_torch/moe/test_moe_load_balancer.py |
@@ -1029,7 +1027,6 @@ legacy-files: &legacy_files |
         tensorrt_llm/_torch/modules/triton_linear.py |
         tensorrt_llm/_torch/moe/expert_statistic.py |
         tensorrt_llm/_torch/moe/fused_moe/__init__.py |
-        tensorrt_llm/_torch/moe/fused_moe/communication/moe_alltoall.py |
         tensorrt_llm/_torch/moe/fused_moe/create_moe.py |
         tensorrt_llm/_torch/moe/fused_moe/deep_ep_utils.py |
         tensorrt_llm/_torch/moe/fused_moe/fused_moe_cute_dsl.py |
@@ -1338,7 +1335,6 @@ legacy-files: &legacy_files |
         tests/unittest/_torch/modeling/test_modeling_vila.py |
         tests/unittest/_torch/modules/test_group_rmn_norm.py |
         tests/unittest/_torch/modules/test_triton_linear.py |
-        tests/unittest/_torch/moe/multi_gpu/test_moe_a2a.py |
         tests/unittest/_torch/moe/test_fused_moe.py |
         tests/unittest/_torch/moe/test_moe_host_sharer.py |
         tests/unittest/_torch/moe/test_moe_load_balancer.py |

cpp/tensorrt_llm/common/envUtils.cpp

@@ -542,50 +542,6 @@ bool getEnvDisableChunkedAttentionInGenPhase()
     return getBoolEnv("TRTLLM_DISABLE_CHUNKED_ATTENTION_IN_GEN_PHASE");
 }
 
-static int sanitizeBlockSize(std::optional<int32_t> const& val)
-{
-    // Default 256 when not set or invalid
-    int block = val.value_or(256);
-    // Clamp to sane CUDA bounds and warp multiples
-    if (block <= 0)
-        block = 256;
-    if (block > 1024)
-        block = 1024;
-    // Round to nearest multiple of 32 (warp size)
-    block = (block + 31) / 32 * 32;
-    if (block == 0)
-        block = 256;
-    return block;
-}
-
-// Read an integer env var and sanitize it as a CUDA block size. Treats malformed
-// values (e.g. non-numeric strings that would throw inside std::stoi) as unset and
-// falls back to the default, so this debug knob never becomes a hard failure.
-static int getSanitizedBlockSizeFromEnv(char const* name)
-{
-    try
-    {
-        return sanitizeBlockSize(getIntEnv(name));
-    }
-    catch (std::exception const&)
-    {
-        TLLM_LOG_WARNING("Invalid value for %s. Falling back to default block size.", name);
-        return sanitizeBlockSize(std::nullopt);
-    }
-}
-
-int getEnvMoeA2ADispatchBlockSize()
-{
-    static int const kBlock = getSanitizedBlockSizeFromEnv("TLLM_MOE_A2A_DISPATCH_BLOCK_SIZE");
-    return kBlock;
-}
-
-int getEnvMoeA2ACombineBlockSize()
-{
-    static int const kBlock = getSanitizedBlockSizeFromEnv("TLLM_MOE_A2A_COMBINE_BLOCK_SIZE");
-    return kBlock;
-}
-
 bool getEnvEplbForceGdrcopy()
 {
     return getBoolEnv("TRTLLM_EPLB_FORCE_GDRCOPY");

cpp/tensorrt_llm/common/envUtils.h

@@ -170,12 +170,6 @@ bool getEnvDisaggBenchmarkGenOnly();
 // Whether to disable the chunked-attention in the generation phase.
 bool getEnvDisableChunkedAttentionInGenPhase();
 
-// TODO: For DEV purpose temporarily.
-// Block size (threads per block) for MoE A2A Dispatch kernels (default 256 if unset or invalid)
-int getEnvMoeA2ADispatchBlockSize();
-// Block size (threads per block) for MoE A2A Combine kernels (default 256 if unset or invalid)
-int getEnvMoeA2ACombineBlockSize();
-
 bool getEnvKVCacheTransferAllBlocksForWindow();
 
 bool getEnvEplbForceGdrcopy();