Skip to content

Commit 6bef27b

Browse files
appleinskyappleinsky
authored
Merge pull request #10 from 1190201122/fix-conv1d-prepack
Fix conv1d prepack（slot_mapping / conv1d_prepack / mm_ar_rmsnorm）
2 parents b1f9676 + 7a6c863 commit 6bef27b

17 files changed

Lines changed: 1084 additions & 12 deletions

‎start_qwen3.6-27b_aclnn_graph.sh‎

Lines changed: 36 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,36 @@
1+
#!/bin/bash
2+
# New test: aclnn chunk_gated_delta_rule, graph mode
3+
4+
export VLLM_FL_USE_ACLNN_CHUNK_GDN=1
5+
# 3 features: all ON
6+
export VLLM_FL_DISABLE_CONV1D_PREPACK=0
7+
export VLLM_FL_ENABLE_MM_AR_RMSNORM=1
8+
export VLLM_FL_DISABLE_NPU_SLOT_MAPPING=0
9+
10+
echo "=========================================="
11+
echo " Running with aclnn chunk_gated_delta_rule, graph mode"
12+
echo "=========================================="
13+
echo "VLLM_FL_USE_ACLNN_CHUNK_GDN=${VLLM_FL_USE_ACLNN_CHUNK_GDN}"
14+
echo "VLLM_FL_DISABLE_CONV1D_PREPACK=${VLLM_FL_DISABLE_CONV1D_PREPACK}"
15+
echo "VLLM_FL_ENABLE_MM_AR_RMSNORM=${VLLM_FL_ENABLE_MM_AR_RMSNORM}"
16+
echo "VLLM_FL_DISABLE_NPU_SLOT_MAPPING=${VLLM_FL_DISABLE_NPU_SLOT_MAPPING}"
17+
echo ""
18+
19+
/workspace/scripts/run_vllm_fl_profile_unified.sh \
20+
--model-path /models/Qwen3.6-27B \
21+
--model-name qwen3.6 \
22+
--model-tag qwen3.6-27b \
23+
--mode graph \
24+
--cudagraph-mode FULL \
25+
--cases "1024,1024,256;4096,1024,256;16384,1024,256;65536,1024,256" \
26+
--concurrency 64 \
27+
--max-num-seqs 64 \
28+
--max-model-len 131072 \
29+
--tp 4 \
30+
--gmem 0.9 \
31+
--devices 0,1,2,3 \
32+
--port 8122 \
33+
--no-bench-profile \
34+
--skip-analyse \
35+
--package none \
36+
--run-label aclnn_chunk_3feat_on_c64_4case_i1k_4k_16k_64k_graph_noprof

‎start_qwen3.6-35b_aclnn_graph.sh‎

Lines changed: 36 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,36 @@
1+
#!/bin/bash
2+
# New test: aclnn chunk_gated_delta_rule, graph mode (35B-A3B)
3+
4+
export VLLM_FL_USE_ACLNN_CHUNK_GDN=1
5+
# features: slot_mapping ON only (conv1d_prepack OFF, mm_ar_rmsnorm OFF)
6+
export VLLM_FL_DISABLE_CONV1D_PREPACK=1
7+
export VLLM_FL_ENABLE_MM_AR_RMSNORM=0
8+
export VLLM_FL_DISABLE_NPU_SLOT_MAPPING=0
9+
10+
echo "=========================================="
11+
echo " Running with aclnn chunk_gated_delta_rule, graph mode (35B-A3B)"
12+
echo "=========================================="
13+
echo "VLLM_FL_USE_ACLNN_CHUNK_GDN=${VLLM_FL_USE_ACLNN_CHUNK_GDN}"
14+
echo "VLLM_FL_DISABLE_CONV1D_PREPACK=${VLLM_FL_DISABLE_CONV1D_PREPACK}"
15+
echo "VLLM_FL_ENABLE_MM_AR_RMSNORM=${VLLM_FL_ENABLE_MM_AR_RMSNORM}"
16+
echo "VLLM_FL_DISABLE_NPU_SLOT_MAPPING=${VLLM_FL_DISABLE_NPU_SLOT_MAPPING}"
17+
echo ""
18+
19+
/workspace/scripts/run_vllm_fl_profile_unified.sh \
20+
--model-path /models/Qwen3.6-35B-A3B \
21+
--model-name qwen3.6 \
22+
--model-tag qwen3.6-35b-a3b \
23+
--mode graph \
24+
--cudagraph-mode FULL \
25+
--cases "1024,1024,256;4096,1024,256;16384,1024,256;65536,1024,256" \
26+
--concurrency 64 \
27+
--max-num-seqs 64 \
28+
--max-model-len 131072 \
29+
--tp 4 \
30+
--gmem 0.9 \
31+
--devices 4,5,6,7 \
32+
--port 8123 \
33+
--no-bench-profile \
34+
--skip-analyse \
35+
--package none \
36+
--run-label slot_on_c64_4case_i1k_4k_16k_64k_graph_noprof

‎start_qwen3.6-35b_aclnn_graph_1k.sh‎

Lines changed: 29 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,29 @@
1+
#!/bin/bash
2+
# New test: aclnn chunk_gated_delta_rule, graph mode (35B-A3B, 1k case only)
3+
4+
export VLLM_FL_USE_ACLNN_CHUNK_GDN=1
5+
6+
echo "=========================================="
7+
echo " Running with aclnn chunk_gated_delta_rule, graph mode (35B-A3B, 1k case)"
8+
echo "=========================================="
9+
echo "VLLM_FL_USE_ACLNN_CHUNK_GDN=${VLLM_FL_USE_ACLNN_CHUNK_GDN}"
10+
echo ""
11+
12+
/workspace/scripts/run_vllm_fl_profile_unified.sh \
13+
--model-path /models/Qwen3.6-35B-A3B \
14+
--model-name qwen3.6 \
15+
--model-tag qwen3.6-35b-a3b \
16+
--mode graph \
17+
--cudagraph-mode FULL \
18+
--cases "1024,1024,256" \
19+
--concurrency 64 \
20+
--max-num-seqs 64 \
21+
--max-model-len 131072 \
22+
--tp 4 \
23+
--gmem 0.9 \
24+
--devices 4,5,6,7 \
25+
--port 8123 \
26+
--no-bench-profile \
27+
--skip-analyse \
28+
--package none \
29+
--run-label aclnn_chunk_c64_i1024_o1024_np256_gpu47_graph_noprof

‎start_qwen3.6-35b_aclnn_graph_1k_gpu0-3.sh‎

Lines changed: 29 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,29 @@
1+
#!/bin/bash
2+
# New test: aclnn chunk_gated_delta_rule, graph mode (35B-A3B, 1k case, gpu0-3)
3+
4+
export VLLM_FL_USE_ACLNN_CHUNK_GDN=1
5+
6+
echo "=========================================="
7+
echo " Running with aclnn chunk_gated_delta_rule, graph mode (35B-A3B, 1k case, gpu0-3)"
8+
echo "=========================================="
9+
echo "VLLM_FL_USE_ACLNN_CHUNK_GDN=${VLLM_FL_USE_ACLNN_CHUNK_GDN}"
10+
echo ""
11+
12+
/workspace/scripts/run_vllm_fl_profile_unified.sh \
13+
--model-path /models/Qwen3.6-35B-A3B \
14+
--model-name qwen3.6 \
15+
--model-tag qwen3.6-35b-a3b \
16+
--mode graph \
17+
--cudagraph-mode FULL \
18+
--cases "1024,1024,256" \
19+
--concurrency 64 \
20+
--max-num-seqs 64 \
21+
--max-model-len 131072 \
22+
--tp 4 \
23+
--gmem 0.9 \
24+
--devices 0,1,2,3 \
25+
--port 8122 \
26+
--no-bench-profile \
27+
--skip-analyse \
28+
--package none \
29+
--run-label aclnn_chunk_c64_i1024_o1024_np256_gpu03_graph_noprof

‎start_qwen3.6-35b_aclnn_graph_3feat_off.sh‎

Lines changed: 36 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,36 @@
1+
#!/bin/bash
2+
# aclnn chunk_gated_delta_rule, graph mode (35B-A3B), 3 features all OFF
3+
4+
export VLLM_FL_USE_ACLNN_CHUNK_GDN=1
5+
# 3 features: all OFF
6+
export VLLM_FL_DISABLE_CONV1D_PREPACK=1
7+
export VLLM_FL_ENABLE_MM_AR_RMSNORM=0
8+
export VLLM_FL_DISABLE_NPU_SLOT_MAPPING=1
9+
10+
echo "=========================================="
11+
echo " Running with aclnn chunk_gated_delta_rule, graph mode (35B-A3B, 3feat OFF)"
12+
echo "=========================================="
13+
echo "VLLM_FL_USE_ACLNN_CHUNK_GDN=${VLLM_FL_USE_ACLNN_CHUNK_GDN}"
14+
echo "VLLM_FL_DISABLE_CONV1D_PREPACK=${VLLM_FL_DISABLE_CONV1D_PREPACK}"
15+
echo "VLLM_FL_ENABLE_MM_AR_RMSNORM=${VLLM_FL_ENABLE_MM_AR_RMSNORM}"
16+
echo "VLLM_FL_DISABLE_NPU_SLOT_MAPPING=${VLLM_FL_DISABLE_NPU_SLOT_MAPPING}"
17+
echo ""
18+
19+
/workspace/scripts/run_vllm_fl_profile_unified.sh \
20+
--model-path /models/Qwen3.6-35B-A3B \
21+
--model-name qwen3.6 \
22+
--model-tag qwen3.6-35b-a3b \
23+
--mode graph \
24+
--cudagraph-mode FULL \
25+
--cases "1024,1024,256;4096,1024,256;16384,1024,256;65536,1024,256" \
26+
--concurrency 64 \
27+
--max-num-seqs 64 \
28+
--max-model-len 131072 \
29+
--tp 4 \
30+
--gmem 0.9 \
31+
--devices 0,1,2,3 \
32+
--port 8122 \
33+
--no-bench-profile \
34+
--skip-analyse \
35+
--package none \
36+
--run-label aclnn_chunk_3feat_off_c64_4case_i1k_4k_16k_64k_graph_noprof

‎start_qwen3.6-35b_baseline_graph.sh‎

Lines changed: 29 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,29 @@
1+
#!/bin/bash
2+
# Baseline test: PTO + Triton (current implementation), graph mode (35B-A3B)
3+
4+
export VLLM_FL_USE_ACLNN_CHUNK_GDN=0
5+
6+
echo "=========================================="
7+
echo " Running Baseline (PTO + Triton), graph mode (35B-A3B)"
8+
echo "=========================================="
9+
echo "VLLM_FL_USE_ACLNN_CHUNK_GDN=${VLLM_FL_USE_ACLNN_CHUNK_GDN}"
10+
echo ""
11+
12+
/workspace/scripts/run_vllm_fl_profile_unified.sh \
13+
--model-path /models/Qwen3.6-35B-A3B \
14+
--model-name qwen3.6 \
15+
--model-tag qwen3.6-35b-a3b \
16+
--mode graph \
17+
--cudagraph-mode FULL \
18+
--cases "1024,1024,256" \
19+
--concurrency 64 \
20+
--max-num-seqs 64 \
21+
--max-model-len 131072 \
22+
--tp 4 \
23+
--gmem 0.9 \
24+
--devices 4,5,6,7 \
25+
--port 8123 \
26+
--no-bench-profile \
27+
--skip-analyse \
28+
--package none \
29+
--run-label baseline_pto_c64_i1024_o1024_np256_gpu47_graph_noprof

‎start_qwen3.6-35b_baseline_graph_gpu0-3.sh‎

Lines changed: 29 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,29 @@
1+
#!/bin/bash
2+
# Baseline test: PTO + Triton (current implementation), graph mode (35B-A3B, gpu0-3)
3+
4+
export VLLM_FL_USE_ACLNN_CHUNK_GDN=0
5+
6+
echo "=========================================="
7+
echo " Running Baseline (PTO + Triton), graph mode (35B-A3B, gpu0-3)"
8+
echo "=========================================="
9+
echo "VLLM_FL_USE_ACLNN_CHUNK_GDN=${VLLM_FL_USE_ACLNN_CHUNK_GDN}"
10+
echo ""
11+
12+
/workspace/scripts/run_vllm_fl_profile_unified.sh \
13+
--model-path /models/Qwen3.6-35B-A3B \
14+
--model-name qwen3.6 \
15+
--model-tag qwen3.6-35b-a3b \
16+
--mode graph \
17+
--cudagraph-mode FULL \
18+
--cases "1024,1024,256" \
19+
--concurrency 64 \
20+
--max-num-seqs 64 \
21+
--max-model-len 131072 \
22+
--tp 4 \
23+
--gmem 0.9 \
24+
--devices 0,1,2,3 \
25+
--port 8122 \
26+
--no-bench-profile \
27+
--skip-analyse \
28+
--package none \
29+
--run-label baseline_pto_c64_i1024_o1024_np256_gpu03_graph_noprof

‎start_qwen3.6-35b_slot_conv1d_on.sh‎

Lines changed: 35 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,35 @@
1+
#!/bin/bash
2+
# slot_mapping ON + conv1d_prepack ON (35B-A3B, gpu0-3)
3+
4+
export VLLM_FL_USE_ACLNN_CHUNK_GDN=1
5+
# features: slot_mapping ON, conv1d_prepack ON, mm_ar_rmsnorm OFF
6+
export VLLM_FL_DISABLE_CONV1D_PREPACK=0
7+
export VLLM_FL_ENABLE_MM_AR_RMSNORM=0
8+
export VLLM_FL_DISABLE_NPU_SLOT_MAPPING=0
9+
10+
echo "=========================================="
11+
echo " slot_mapping ON + conv1d_prepack ON (35B-A3B)"
12+
echo "=========================================="
13+
echo "VLLM_FL_DISABLE_CONV1D_PREPACK=${VLLM_FL_DISABLE_CONV1D_PREPACK}"
14+
echo "VLLM_FL_ENABLE_MM_AR_RMSNORM=${VLLM_FL_ENABLE_MM_AR_RMSNORM}"
15+
echo "VLLM_FL_DISABLE_NPU_SLOT_MAPPING=${VLLM_FL_DISABLE_NPU_SLOT_MAPPING}"
16+
echo ""
17+
18+
/workspace/scripts/run_vllm_fl_profile_unified.sh \
19+
--model-path /models/Qwen3.6-35B-A3B \
20+
--model-name qwen3.6 \
21+
--model-tag qwen3.6-35b-a3b \
22+
--mode graph \
23+
--cudagraph-mode FULL \
24+
--cases "1024,1024,256;4096,1024,256;16384,1024,256;65536,1024,256" \
25+
--concurrency 64 \
26+
--max-num-seqs 64 \
27+
--max-model-len 131072 \
28+
--tp 4 \
29+
--gmem 0.9 \
30+
--devices 0,1,2,3 \
31+
--port 8122 \
32+
--no-bench-profile \
33+
--skip-analyse \
34+
--package none \
35+
--run-label slot_conv1d_on_c64_4case_i1k_4k_16k_64k_graph_noprof

‎start_qwen3.6-35b_slot_mmar_on.sh‎

Lines changed: 41 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,41 @@
1+
#!/bin/bash
2+
# slot_mapping ON + mm_ar_rmsnorm ON (35B-A3B, gpu4-7)
3+
# mm fusion threshold: M > 512 (VLLM_FL_MM_AR_RMSNORM_MIN_TOKENS).
4+
# Rationale: decode steps have M <= concurrency (64) and must NOT fuse
5+
# (the extra add_out all-gather cancels the win at small M); prefill chunks
6+
# (up to 2048) fuse only above 512, same as the tuned npugraph_ex pass.
7+
8+
export VLLM_FL_USE_ACLNN_CHUNK_GDN=1
9+
# features: slot_mapping ON, conv1d_prepack OFF, mm_ar_rmsnorm ON
10+
export VLLM_FL_DISABLE_CONV1D_PREPACK=1
11+
export VLLM_FL_ENABLE_MM_AR_RMSNORM=1
12+
export VLLM_FL_DISABLE_NPU_SLOT_MAPPING=0
13+
export VLLM_FL_MM_AR_RMSNORM_MIN_TOKENS=512
14+
15+
echo "=========================================="
16+
echo " slot_mapping ON + mm_ar_rmsnorm ON (35B-A3B)"
17+
echo "=========================================="
18+
echo "VLLM_FL_DISABLE_CONV1D_PREPACK=${VLLM_FL_DISABLE_CONV1D_PREPACK}"
19+
echo "VLLM_FL_ENABLE_MM_AR_RMSNORM=${VLLM_FL_ENABLE_MM_AR_RMSNORM}"
20+
echo "VLLM_FL_DISABLE_NPU_SLOT_MAPPING=${VLLM_FL_DISABLE_NPU_SLOT_MAPPING}"
21+
echo "VLLM_FL_MM_AR_RMSNORM_MIN_TOKENS=${VLLM_FL_MM_AR_RMSNORM_MIN_TOKENS}"
22+
echo ""
23+
24+
/workspace/scripts/run_vllm_fl_profile_unified.sh \
25+
--model-path /models/Qwen3.6-35B-A3B \
26+
--model-name qwen3.6 \
27+
--model-tag qwen3.6-35b-a3b \
28+
--mode graph \
29+
--cudagraph-mode FULL \
30+
--cases "1024,1024,256;4096,1024,256;16384,1024,256;65536,1024,256" \
31+
--concurrency 64 \
32+
--max-num-seqs 64 \
33+
--max-model-len 131072 \
34+
--tp 4 \
35+
--gmem 0.9 \
36+
--devices 4,5,6,7 \
37+
--port 8123 \
38+
--no-bench-profile \
39+
--skip-analyse \
40+
--package none \
41+
--run-label slot_mmar_on_c64_4case_i1k_4k_16k_64k_graph_noprof

0 commit comments

Comments
 (0)