sunxin adc69c15d0 [Model Runner] Prepare token count and move FA3 initialization into the graph (#6170) ...

* prepare for token num and put FA3 init in graph

2026-01-26 12:16:57 +08:00

..

template_instantiation

…

append_attention_c4_impl.cuh

[Optimization] Do not compute ATTN padding part in In Cuda graph mode (#5985)

2026-01-13 11:32:27 +08:00

append_attention_c8_impl.cuh

[Optimization] Do not compute ATTN padding part in In Cuda graph mode (#5985)

2026-01-13 11:32:27 +08:00

append_attention_c16_impl.cuh

[Optimization] Do not compute ATTN padding part in In Cuda graph mode (#5985)

2026-01-13 11:32:27 +08:00

append_attention_func.cuh

[Model Runner] Prepare token count and move FA3 initialization into the graph (#6170)

2026-01-26 12:16:57 +08:00

append_attention_kernel.h

…

decode_attention_func.cuh

…

decoder_mla_attention_kernel.cu

…

decoder_mla_attention_kernel.h

…

decoder_write_cache_with_rope_impl.cuh

[Feature][Optimization] Qwen Support Dynamic block_wise_fp8 cache (#5486)

2025-12-12 17:10:17 +08:00

decoder_write_cache_with_rope_kernel.cu

[Feature][Optimization] Qwen Support Dynamic block_wise_fp8 cache (#5486)

2025-12-12 17:10:17 +08:00

decoder_write_cache_with_rope_kernel.h

…

encoder_write_cache_with_rope_impl.cuh

[Model Runner] Prepare token count and move FA3 initialization into the graph (#6170)

2026-01-26 12:16:57 +08:00

encoder_write_cache_with_rope_kernel.h

…

get_block_shape_and_split_kv_block.cu

[Optimization] Do not compute ATTN padding part in In Cuda graph mode (#5985)

2026-01-13 11:32:27 +08:00

gqa_rope_write_cache.cu

[Model Runner] Prepare token count and move FA3 initialization into the graph (#6170)

2026-01-26 12:16:57 +08:00

mem_util.cuh

…

mla_cache_kernel.cu

[PD Disaggregation] Support PD deployment of DeepSeekv3. (#5251)

2025-12-02 14:11:50 +08:00

mla_cache_kernel.cuh

…

mma_tensor_op.cuh

…

multiquery_attention_c4_impl.cuh

[Optimization] Do not compute ATTN padding part in In Cuda graph mode (#5985)

2026-01-13 11:32:27 +08:00

multiquery_attention_c4_kernel.h

…

multiquery_attention_c8_impl.cuh

[Optimization] Do not compute ATTN padding part in In Cuda graph mode (#5985)

2026-01-13 11:32:27 +08:00

multiquery_attention_c8_kernel.h

…

multiquery_attention_c16_impl.cuh

[Optimization] Do not compute ATTN padding part in In Cuda graph mode (#5985)

2026-01-13 11:32:27 +08:00

multiquery_attention_c16_kernel.h

…

multiquery_decoder_attention_impl.cuh

…

multiquery_decoder_attention_kernel.h

…

pre_cache_len_concat.cu

[Others] Remove useless code (#5404)

2025-12-08 13:59:46 +08:00

qwen3_rope.h

support fa3 qwen-vl rope (#5869)

2026-01-05 15:29:34 +08:00

speculate_write_cache_with_rope_impl.cuh

[Speculative Decoding] Support MTP for GLM-4.5-Air (#6047)

2026-01-16 14:35:24 +08:00

speculate_write_cache_with_rope_kernel.cu

[Speculative Decoding] Support MTP for GLM-4.5-Air (#6047)

2026-01-16 14:35:24 +08:00

speculate_write_cache_with_rope_kernel.h

…

template_config.json

revert group size 3 (#5079)

2025-11-17 18:54:13 +08:00

utils.cuh

…