Skip to content

Port TRT-LLM communication kernels to flashinfer #2102

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
wants to merge 9 commits into
base: main
Choose a base branch
from
Draft

Port TRT-LLM communication kernels to flashinfer #2102

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
2d9d327
Initial draft of cursor port
djns99 Nov 18, 2025
0974614
Properly cache jit compiled module
djns99 Nov 19, 2025
0ebd5cd
Cleanup prints
djns99 Nov 19, 2025
25074e6
Combine tests run without crashing
djns99 Nov 19, 2025
8220b9e
Update tests with fake_moe properly
djns99 Nov 19, 2025
fede35b
Clear MOE workspace before each run
djns99 Nov 19, 2025
2bac65a
Cleanup MPI processes on test failures
djns99 Nov 20, 2025
bd82a2b
More exit handling for rank failures
djns99 Nov 20, 2025
c2a099c
Cleaner test implementation
djns99 Nov 20, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
40 changes: 34 additions & 6 deletions csrc/nv_internal/cpp/common/envUtils.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -222,11 +222,6 @@ bool getEnvDisaggLayerwise() {
return disaggLayerwise;
}

bool getEnvParallelCacheSend() {
static bool const parallelCacheSend = getBoolEnv("TRTLLM_PARALLEL_CACHE_SEND");
return parallelCacheSend;
}

bool getEnvRequestKVCacheConcurrent() {
static bool const requestKVCacheConcurrent = getBoolEnv("TRTLLM_REQUEST_KV_CACHE_CONCURRENT");
return requestKVCacheConcurrent;
Expand Down Expand Up @@ -277,7 +272,7 @@ size_t getEnvAllReduceWorkspaceSize() {
return workspaceSize;
}

std::string getEnvKVCacheTransferOutputPath() {
std::string const& getEnvKVCacheTimeOutputPath() {
static std::string outputPath = getStrEnv("TRTLLM_KVCACHE_TIME_OUTPUT_PATH").value_or("");
return outputPath;
}
Expand Down Expand Up @@ -328,4 +323,37 @@ uint16_t getEnvNixlPort() {

bool getEnvDisaggBenchmarkGenOnly() { return getBoolEnv("TRTLLM_DISAGG_BENCHMARK_GEN_ONLY"); }

bool getEnvMoeA2AOneBlockPerToken() {
// Default true; return false only if env set to "0"
static std::optional<int32_t> const val = getIntEnv("TLLM_MOE_A2A_ONE_BLOCK_PER_TOKEN");
if (!val.has_value()) {
return true;
}
return val.value() != 0;
}

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;
}

int getEnvMoeA2ADispatchBlockSize() {
static int const kBlock = sanitizeBlockSize(getIntEnv("TLLM_MOE_A2A_DISPATCH_BLOCK_SIZE"));
return kBlock;
}

int getEnvMoeA2ACombineBlockSize() {
static int const kBlock = sanitizeBlockSize(getIntEnv("TLLM_MOE_A2A_COMBINE_BLOCK_SIZE"));
return kBlock;
}

bool getEnvEplbForceGdrcopy() { return getBoolEnv("TRTLLM_EPLB_FORCE_GDRCOPY"); }

} // namespace tensorrt_llm::common
11 changes: 10 additions & 1 deletion csrc/nv_internal/tensorrt_llm/common/envUtils.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -64,7 +64,7 @@ bool getEnvDisableKVCacheTransferOverlap();

bool getEnvEnableReceiveKVCacheParallel();

std::string getEnvKVCacheTransferOutputPath();
std::string const& getEnvKVCacheTimeOutputPath();

bool getEnvTryZCopyForKVCacheTransfer();

Expand Down Expand Up @@ -92,4 +92,13 @@ size_t getEnvKVCacheSendMaxConcurrenceNum();

size_t getEnvMemSizeForKVCacheTransferBuffer();

// Whether to use one block per token for MoE A2A kernels (default true).
bool getEnvMoeA2AOneBlockPerToken();

// 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();

} // namespace tensorrt_llm::common
Loading