Jundong Liu 1e4968e810 [Excutor] Fixed the issue of CUDA graph execution failure caused by different branches during decoding (#3223) ...

* 彻底解决解码切块问题

* update C8 and C4 kernel

* fix problem

* fix with pre-commit

* retain branch for mtp

2025-08-09 07:37:19 +08:00

..

append_attn

[Excutor] Fixed the issue of CUDA graph execution failure caused by different branches during decoding (#3223)

2025-08-09 07:37:19 +08:00

common

…

custom_all_reduce

…

cutlass_extensions

Optimize the performance of moe_expert_ffn_wint2 (#2990)

2025-07-28 10:32:43 +08:00

cutlass_kernels

optimize w4a8 decoding (#3050)

2025-07-28 22:20:13 +08:00

flash_mask_attn

[New Feature] fa3 支持flash mask (#3184)

2025-08-05 12:20:48 +08:00

fp8_gemm_with_cutlass

…

glog

…

int8_gemm_with_cutlass

…

mla_attn

[BugFix] Rename attention params of deepseekv3 (#2939)

2025-07-22 14:01:30 +08:00

moe

support qk norm (#3145)

2025-08-05 16:46:14 +08:00

quantization

…

sample_kernels

[Iluvatar GPU] Optimze attention and moe performance (#3234)

2025-08-08 10:51:24 +08:00

speculate_decoding

[Feature] support bad_words (#3055)

2025-07-30 09:31:29 +08:00

w4afp8_gemm

[New Feature] Support W4Afp8 MoE GroupGemm (#3171)

2025-08-06 10:34:05 +08:00

append_attention.cu

support qk norm (#3145)

2025-08-05 16:46:14 +08:00

beam_search_softmax.cu

…

cpp_extensions.cc

support qk norm (#3145)

2025-08-05 16:46:14 +08:00

cuda_multiprocess.h

…

dequant_int8.cu

…

enforce_generation.cu

…

env.h

…

extract_text_token_output.cu

[Code Simplification] Refactor Post-processing in VL Model Forward Method (#2937)

2025-08-01 17:28:07 +08:00

fused_get_rope.cu

…

fused_hadamard_quant_fp8.cu

…

fused_rotary_position_encoding.cu

[Feature] DeepseekV3 use pd_build_static_op (#2948)

2025-07-22 15:03:41 +08:00

gather_idx.cu

…

get_data_ptr_ipc.cu

…

get_img_boundaries.cc

[Feature] Multimodal Scheduler V1 (#3019)

2025-07-30 16:05:55 +08:00

get_mm_split_fuse.cc

…

get_output_ep.cc

…

get_output_msg_with_topk.cc

…

get_output.cc

…

get_padding_offset_system.cu

…

get_padding_offset.cu

…

get_position_ids_and_mask_encoder_batch.cu

[Feature] DeepseekV3 use pd_build_static_op (#2948)

2025-07-22 15:03:41 +08:00

helper.cu

…

helper.h

…

init_signal_layerwise.cc

…

ipc_sent_key_value_cache_by_remote_ptr.cu

…

merge_prefill_decode_output.cu

集中式支持fa3 (#3112)

2025-08-01 18:03:36 +08:00

msg_utils.h

…

multi_head_latent_attention.cu

[Feature] DeepseekV3 use pd_build_static_op (#2948)

2025-07-22 15:03:41 +08:00

ngram_mask.cu

…

noaux_tc.cu

[Bugfix] Fix model accuracy in some ops (#3231)

2025-08-08 17:30:37 +08:00

noauxtc_kernel.h

revise get_moe_scores (#3164)

2025-08-05 16:43:07 +08:00

open_shm_and_get_meta_signal.cc

…

per_token_quant_fp8.cu

Add switch to apply fine-grained per token quant fp8 (#3192)

2025-08-04 19:54:03 -07:00

read_data_ipc.cu

…

read_ids.py

…

read_temp_ids.py

…

rebuild_padding.cu

…

recover_decode_task.cu

[Feature] Support block scheduler v1 for FD (#2928)

2025-07-23 20:31:31 +08:00

remote_cache_kv_ipc.cc

…

remote_cache_kv_ipc.h

…

save_output_msg_with_topk.cc

…

save_with_output_msg.cc

…

save_with_output_msg.h

…

save_with_output.cc

…

scaled_gemm_f8_i4_f16_gemm.cu

…

scaled_gemm_f8_i4_f16_weight_quantize.cu

…

seqs2seqs.cu

…

set_data_ipc.cu

…

set_flags.cu

…

set_mask_value.cu

…

set_value_by_flags.cu

…

share_external_data.cu

…

step_reschedule.cu

…

step_system_cache.cu

…

step.cu

…

stop_generation_multi_ends.cu

[stop sequence] support stop sequence (#3025)

2025-07-29 14:17:37 +08:00

stop_generation.cu

…

swap_cache_batch.cu

…

swap_cache.cu

…

system2group.cu

…

text_image_gather_scatter.cu

…

text_image_index_out.cu

…

token_penalty_multi_scores.cu

[Feature] support bad_words (#3055)

2025-07-30 09:31:29 +08:00

token_penalty_only_once.cu

…

token_transfer.hpp

…

transfer_output.cc

…

tune_cublaslt_gemm.cu

…

update_inputs_beam.cu

…

update_inputs_v1.cu

[Feature] Support block scheduler v1 for FD (#2928)

2025-07-23 20:31:31 +08:00

update_inputs.cu

…

update_split_fuse_input.cu

…