Skip to content

Update the routing for TRTLLMGEN to support kimi k2 and qwen #1831

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

Merged
merged 10 commits into from
Oct 22, 2025
+1,036 −711
Merged

Update the routing for TRTLLMGEN to support kimi k2 and qwen #1831

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
49aedec
Update the routing for TRTLLMGEN to support kimi k2 and qwen
ChristinaZ Oct 1, 2025
db83e76
update conftest
jiahanc Oct 16, 2025
cbd4150
fix wrong size calculation
jiahanc Oct 18, 2025
3d2a5e8
apply fix to fp8 pt
jiahanc Oct 18, 2025
24d0eaf
fix wrong cache name and remove some commented code
jiahanc Oct 18, 2025
1dd6ad9
lint and revert some change
jiahanc Oct 18, 2025
dba345c
add more value to cache key
jiahanc Oct 18, 2025
5cfa98b
fix test
jiahanc Oct 19, 2025
63233a3
fix merge conflict
jiahanc Oct 20, 2025
4b38452
remove epilogue and sf in cache key, not need
jiahanc Oct 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
138 changes: 87 additions & 51 deletions csrc/trtllm_fused_moe_kernel_launcher.cu
View file
Open in desktop

Large diffs are not rendered by default.

418 changes: 253 additions & 165 deletions csrc/trtllm_fused_moe_routing_deepseek.cu
View file
Open in desktop

Large diffs are not rendered by default.

227 changes: 143 additions & 84 deletions csrc/trtllm_fused_moe_routing_llama4.cu
View file
Open in desktop

Large diffs are not rendered by default.

322 changes: 239 additions & 83 deletions csrc/trtllm_fused_moe_routing_renormalize.cu
View file
Open in desktop

Large diffs are not rendered by default.

12 changes: 6 additions & 6 deletions csrc/trtllm_fused_moe_runner.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -70,12 +70,12 @@ void Runner::run(void* routingLogits, void* routingBias, int32_t numTokens, int3
routingData.mUsePdl = true;

// output:
routingData.mPtrExpertIdx = routingExpertIndexes;
routingData.mPtrTopKPacked = routingExpertIndexes;
routingData.mPtrExpertCounts = expertCountHistogram;
routingData.mPtrPermutedIdxSize = permutedIdxSize;
routingData.mPtrExpandedIdxToPermutedIdx = expandedIdxToPermutedIdx;
routingData.mPtrPermutedIdxToTokenIdx = permutedIdxToTokenIdx;
routingData.mPtrExpertWeights = expertWeights;
routingData.mPtrTopKWeights = expertWeights;

routingData.mPtrCtaIdxXyToBatchIdx = ctaIdxXyToBatchIdx;
routingData.mPtrCtaIdxXyToMnLimit = ctaIdxXyToMnLimit;
Expand Down Expand Up @@ -107,12 +107,12 @@ void Runner::run(void* routingLogits, void* routingBias, int32_t numTokens, int3
routingData.mUsePdl = true;

// output:
routingData.mPtrExpertIdx = routingExpertIndexes;
routingData.mPtrTopKPacked = routingExpertIndexes;
routingData.mPtrExpertCounts = expertCountHistogram;
routingData.mPtrPermutedIdxSize = permutedIdxSize;
routingData.mPtrExpandedIdxToPermutedIdx = expandedIdxToPermutedIdx;
routingData.mPtrPermutedIdxToTokenIdx = permutedIdxToTokenIdx;
routingData.mPtrExpertWeights = expertWeights;
routingData.mPtrTopKWeights = expertWeights;

routingData.mPtrCtaIdxXyToBatchIdx = ctaIdxXyToBatchIdx;
routingData.mPtrCtaIdxXyToMnLimit = ctaIdxXyToMnLimit;
Expand Down Expand Up @@ -149,12 +149,12 @@ void Runner::run(void* routingLogits, void* routingBias, int32_t numTokens, int3
//
// Outputs
//
routingData.mPtrExpertIdx = routingExpertIndexes;
routingData.mPtrTopKPacked = routingExpertIndexes;
routingData.mPtrExpertCounts = expertCountHistogram;
routingData.mPtrPermutedIdxSize = permutedIdxSize;
routingData.mPtrExpandedIdxToPermutedIdx = expandedIdxToPermutedIdx;
routingData.mPtrPermutedIdxToTokenIdx = permutedIdxToTokenIdx;
routingData.mPtrExpertWeights = expertWeights;
routingData.mPtrTopKWeights = expertWeights;

//
// Grouped Gemm Launch Config Buffers
Expand Down
53 changes: 29 additions & 24 deletions flashinfer/fused_moe/core.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -171,7 +171,9 @@ def _maybe_get_cached_w3_w1_permute_indices(
epilogue_tile_m: int,
num_elts_per_sf: Union[None, int] = None,
) -> torch.Tensor:
if dst_w3_w1_weight.shape not in _cache_permute_indices:
# Create a unique cache key (weight_type, weight_shape)
cache_key = ("w3_w1", dst_w3_w1_weight.shape)
if cache_key not in _cache_permute_indices:
# Get permute indices and chain them together
permute0 = get_reorder_rows_for_gated_act_gemm_row_indices(dst_w3_w1_weight)
if num_elts_per_sf is None:
Expand All @@ -185,10 +187,10 @@ def _maybe_get_cached_w3_w1_permute_indices(
num_elts_per_sf=num_elts_per_sf,
)
# Memoize permute indices as recompute is **very** costly
_cache_permute_indices[dst_w3_w1_weight.shape] = permute0[permute1].to(
_cache_permute_indices[cache_key] = permute0[permute1].to(
dst_w3_w1_weight.device
)
permute_indices = _cache_permute_indices[dst_w3_w1_weight.shape]
permute_indices = _cache_permute_indices[cache_key]
return permute_indices


Expand All @@ -198,7 +200,9 @@ def get_w2_permute_indices_with_cache(
epilogue_tile_m: int,
num_elts_per_sf: Union[None, int] = None,
) -> torch.Tensor:
if dst_w2_weight.shape not in _cache_permute_indices:
# Create a unique cache key (weight_type, weight_shape)
cache_key = ("w2", dst_w2_weight.shape)
if cache_key not in _cache_permute_indices:
if num_elts_per_sf is None:
permute_indices = get_shuffle_matrix_a_row_indices(
dst_w2_weight, epilogue_tile_m
Expand All @@ -210,8 +214,8 @@ def get_w2_permute_indices_with_cache(
num_elts_per_sf=num_elts_per_sf,
).to(dst_w2_weight.device)
# Memoize permute indices as recompute is **very** costly
_cache_permute_indices[dst_w2_weight.shape] = permute_indices
permute_indices = _cache_permute_indices[dst_w2_weight.shape]
_cache_permute_indices[cache_key] = permute_indices
permute_indices = _cache_permute_indices[cache_key]
return permute_indices


Expand Down Expand Up @@ -1073,12 +1077,12 @@ def trtllm_fp8_per_tensor_scale_moe_op(
output2_scales_scalar: torch.Tensor,
num_experts: int,
top_k: int,
n_group: int,
topk_group: int,
n_group: Optional[int],
topk_group: Optional[int],
intermediate_size: int,
local_expert_offset: int,
local_num_experts: int,
routed_scaling_factor: float,
routed_scaling_factor: Optional[float],
use_routing_scales_on_input: bool,
tile_tokens_dim: int = 8,
routing_method_type: int = 0,
Expand Down Expand Up @@ -1127,12 +1131,12 @@ def _fake_trtllm_fp8_per_tensor_scale_moe(
output2_scales_scalar: torch.Tensor,
num_experts: int,
top_k: int,
n_group: int,
topk_group: int,
n_group: Optional[int],
topk_group: Optional[int],
intermediate_size: int,
local_expert_offset: int,
local_num_experts: int,
routed_scaling_factor: float,
routed_scaling_factor: Optional[float],
use_routing_scales_on_input: bool,
tile_tokens_dim: int = 8,
routing_method_type: int = 0,
Expand All @@ -1159,12 +1163,12 @@ def trtllm_fp8_block_scale_moe_op(
output: torch.Tensor,
num_experts: int,
top_k: int,
n_group: int,
topk_group: int,
n_group: Optional[int],
topk_group: Optional[int],
intermediate_size: int,
local_expert_offset: int,
local_num_experts: int,
routed_scaling_factor: float,
routed_scaling_factor: Optional[float],
tile_tokens_dim: int,
routing_method_type: int,
use_shuffled_weight: bool = False,
Expand All @@ -1173,6 +1177,7 @@ def trtllm_fp8_block_scale_moe_op(
) -> torch.Tensor:
if enable_pdl is None:
enable_pdl = device_support_pdl(hidden_states.device)

# Call the C++ function for block scale MoE
moe_op.trtllm_fp8_block_scale_moe(
routing_logits,
Expand Down Expand Up @@ -1214,12 +1219,12 @@ def _fake_trtllm_fp8_block_scale_moe(
output: torch.Tensor,
num_experts: int,
top_k: int,
n_group: int,
topk_group: int,
n_group: Optional[int],
topk_group: Optional[int],
intermediate_size: int,
local_expert_offset: int,
local_num_experts: int,
routed_scaling_factor: float,
routed_scaling_factor: Optional[float],
tile_tokens_dim: int = 8,
routing_method_type: int = 0,
use_shuffled_weight: bool = False,
Expand Down Expand Up @@ -1479,12 +1484,12 @@ def trtllm_fp8_per_tensor_scale_moe(
output2_scales_scalar: torch.Tensor,
num_experts: int,
top_k: int,
n_group: int,
topk_group: int,
n_group: Optional[int],
topk_group: Optional[int],
intermediate_size: int,
local_expert_offset: int,
local_num_experts: int,
routed_scaling_factor: float,
routed_scaling_factor: Optional[float],
use_routing_scales_on_input: bool,
tile_tokens_dim: int = 8,
routing_method_type: int = 0,
Expand Down Expand Up @@ -1552,12 +1557,12 @@ def trtllm_fp8_block_scale_moe(
gemm2_weights_scale: torch.Tensor,
num_experts: int,
top_k: int,
n_group: int,
topk_group: int,
n_group: Optional[int],
topk_group: Optional[int],
intermediate_size: int,
local_expert_offset: int,
local_num_experts: int,
routed_scaling_factor: float,
routed_scaling_factor: Optional[float],
tile_tokens_dim: int = 8,
routing_method_type: int = 0,
use_shuffled_weight: bool = False,
Expand Down
Loading