minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[NPU] fix extend_seq_lens_cpu shape in eager mode (#36843)

Browse Source
This commit is contained in:
heziiop
2026-09-04 09:03:30 +08:00
committed by GitHub
parent 6147a54ddf
commit 9ed2721c6d
2 changed files with 15 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/hardware_backend/npu/attention/ascend_backend.py
+10 -3
View File
@@ -2021,9 +2021,12 @@ class AscendAttnBackend(AttentionBackend):
) )
if forward_batch.forward_mode.is_draft_extend_v2(): if forward_batch.forward_mode.is_draft_extend_v2():
actual_seq_lengths = ( extend_seq_lens_cpu = forward_batch.extend_seq_lens_cpu
np.array(forward_batch.extend_seq_lens_cpu).cumsum().tolist() if not self.graph_mode:
) extend_seq_lens_cpu = extend_seq_lens_cpu[
: forward_batch._original_batch_size
]
actual_seq_lengths = np.array(extend_seq_lens_cpu).cumsum().tolist()
else: else:
actual_seq_lengths = np.arange( actual_seq_lengths = np.arange(
self.speculative_num_draft_tokens, self.speculative_num_draft_tokens,
@@ -2031,6 +2034,10 @@ class AscendAttnBackend(AttentionBackend):
self.speculative_num_draft_tokens, self.speculative_num_draft_tokens,
) )
if not self.graph_mode:
actual_bs = len(actual_seq_lengths)
actual_seq_lengths_kv = actual_seq_lengths_kv[:actual_bs]
is_swa_layer = layer.sliding_window_size != -1 is_swa_layer = layer.sliding_window_size != -1
if ( if (
is_swa_layer is_swa_layer
test/registered/npu/performance/minimax_m2_5/test_npu_minimax_m2_5_w8a8_4p_in64k_out1k_prefix90_50ms.py
+5 -6
View File
@@ -25,7 +25,7 @@ MINIMAX_M2_5_W8A8_4P_IN64K_OUT1K_PREFIX90_ENVS = {
"DEEP_NORMAL_MODE_USE_INT8_QUANT": "1", "DEEP_NORMAL_MODE_USE_INT8_QUANT": "1",
"DEEPEP_HCCL_BUFFSIZE": "1024", "DEEPEP_HCCL_BUFFSIZE": "1024",
"SGLANG_EXTERNAL_MODEL_PACKAGE": "custom_eagle3", "SGLANG_EXTERNAL_MODEL_PACKAGE": "custom_eagle3",
"SGLANG_PREFILL_DELAYER_MAX_PREFILL_BS_WINDOW_SIZE": "128", "SGLANG_PREFILL_DELAYER_MAX_PREFILL_BS_WINDOW_SIZE": "999999999",
"PYTHONPATH": f"{MINIMAX_M2_5_EAGLE3_MODEL_PATH}:{os.environ.get('PYTHONPATH', '')}", "PYTHONPATH": f"{MINIMAX_M2_5_EAGLE3_MODEL_PATH}:{os.environ.get('PYTHONPATH', '')}",
} }
@@ -35,14 +35,14 @@ MINIMAX_M2_5_W8A8_4P_IN64K_OUT1K_PREFIX90_OTHER_ARGS = [
"--mem-fraction-static", "--mem-fraction-static",
0.63, 0.63,
"--max-running-requests", "--max-running-requests",
26, 24,
"--reasoning-parser", "--reasoning-parser",
"minimax-append-think", "minimax-append-think",
"--tool-call-parser", "--tool-call-parser",
"minimax-m2", "minimax-m2",
"--enable-prefill-delayer", "--enable-prefill-delayer",
"--prefill-delayer-max-delay-passes", "--prefill-delayer-max-delay-passes",
128, 30,
"--prefill-max-requests", "--prefill-max-requests",
10, 10,
"--chunked-prefill-size", "--chunked-prefill-size",
@@ -59,7 +59,6 @@ MINIMAX_M2_5_W8A8_4P_IN64K_OUT1K_PREFIX90_OTHER_ARGS = [
20, 20,
22, 22,
24, 24,
26,
"--moe-a2a-backend", "--moe-a2a-backend",
"ascend_fuseep", "ascend_fuseep",
"--fuseep-mode", "--fuseep-mode",
@@ -100,8 +99,8 @@ class TestNPUMiniMaxM2_5W8A8_4P_In64k_Out1k_Prefix90_50ms(
other_args = MINIMAX_M2_5_W8A8_4P_IN64K_OUT1K_PREFIX90_OTHER_ARGS other_args = MINIMAX_M2_5_W8A8_4P_IN64K_OUT1K_PREFIX90_OTHER_ARGS
envs = MINIMAX_M2_5_W8A8_4P_IN64K_OUT1K_PREFIX90_ENVS envs = MINIMAX_M2_5_W8A8_4P_IN64K_OUT1K_PREFIX90_ENVS
dataset_name = "generated-shared-prefix" dataset_name = "generated-shared-prefix"
max_concurrency = 26 max_concurrency = 24
num_prompts = 104 num_prompts = 96
input_len = 65536 input_len = 65536
output_len = 1024 output_len = 1024
random_range_ratio = 1 random_range_ratio = 1