Skip to content

Commit 81469e4

Browse files
authored
Merge pull request #15 from 1190201122/simplified-csrc
simplified-csrc/ascend kernel for qwen3.6 27B & 35B
2 parents b0a9c07 + 3bbad80 commit 81469e4

615 files changed

Lines changed: 17 additions & 161277 deletions

File tree

Some content is hidden

Large Commits have some content hidden by default. Use the searchbox below for content that may be hidden.

csrc/CMakeLists.txt

Lines changed: 1 addition & 11 deletions
Original file line numberDiff line numberDiff line change
@@ -89,21 +89,11 @@ if(VLLM_VENDOR STREQUAL "ascend")
8989
ascend/aclnn_torch_adapter/NPUStorageImpl.h
9090
# Framework operator torch adapters (op_host/op_kernel/ are built by the
9191
# CANN toolchain and are present here for reference / vendor packaging).
92+
ascend/attention/chunk_gated_delta_rule/chunk_gated_delta_rule_torch_adpt.h
9293
ascend/attention/fused_gdn_gating/fused_gdn_gating_torch_adpt.h
93-
ascend/attention/kv_quant_sparse_flash_attention/kv_quant_sparse_flash_attention_torch_adpt.h
94-
ascend/attention/lightning_indexer/lightning_indexer_torch_adpt.h
95-
ascend/attention/lightning_indexer_quant/lightning_indexer_quant_torch_adpt.h
96-
ascend/attention/lightning_indexer_vllm/lightning_indexer_vllm_torch_adpt.h
97-
ascend/attention/ngram_spec_decode/ngram_spec_decode_torch_adpt.h
9894
ascend/attention/recurrent_gated_delta_rule/recurrent_gated_delta_rule_torch_adpt.h
99-
ascend/attention/sparse_flash_attention/sparse_flash_attention_torch_adpt.h
100-
ascend/attention/store_kv_block/store_kv_block_torch_adpt.h
101-
ascend/gmm/grouped_matmul_swiglu_quant_v2/grouped_matmul_swiglu_quant_v2_torch_adpt.h
102-
ascend/gmm/grouped_matmul_swiglu_quant_weight_nz_tensor_list/grouped_matmul_swiglu_quant_torch_adpt.h
103-
ascend/mc2/dispatch_ffn_combine/dispatch_ffn_combine_torch_adpt.h
10495
ascend/mc2/matmul_allreduce_add_rmsnorm/matmul_allreduce_add_rmsnorm_torch_adpt.h
10596
ascend/moe/add_rms_norm_bias/add_rms_norm_bias_torch_adpt.h
106-
ascend/moe/apply_top_k_top_p_custom/apply_top_k_top_p_custom_torch_adpt.h
10797
ascend/moe/moe_gating_top_k/moe_gating_top_k_torch_adpt.h
10898
ascend/moe/moe_init_routing_custom/moe_init_routing_custom_torch_adpt.h
10999
)

csrc/ascend/CMakeLists.txt

Lines changed: 0 additions & 15 deletions
Original file line numberDiff line numberDiff line change
@@ -131,24 +131,9 @@ if(ASCEND950_AVAILABLE)
131131
else()
132132
message(STATUS "ASCEND950 is NOT available in CANN headers; skipping ASCEND950-only operators.")
133133
set(ASCEND950_SKIP_OPS
134-
compressor
135-
inplace_partial_rotary_mul
136-
kv_quant_sparse_flash_attention
137-
lightning_indexer
138134
recurrent_gated_delta_rule
139-
sparse_flash_attention
140-
vllm_quant_lightning_indexer
141135
add_rms_norm_bias
142136
causal_conv1d
143-
dequant_swiglu_quant
144-
hc_post
145-
hc_pre
146-
hc_pre_inv_rms
147-
hc_pre_sinkhorn
148-
moe_gating_top_k
149-
moe_gating_top_k_hash
150-
scatter_nd_update_v2
151-
grouped_matmul_swiglu_quant_v2
152137
)
153138
endif()
154139

csrc/ascend/attention/compressor/CMakeLists.txt

Lines changed: 0 additions & 19 deletions
This file was deleted.

csrc/ascend/attention/compressor/README.md

Lines changed: 0 additions & 499 deletions
This file was deleted.

csrc/ascend/attention/compressor/op_host/CMakeLists.txt

Lines changed: 0 additions & 40 deletions
This file was deleted.

0 commit comments

Comments
 (0)