fix tpot by adjusting the sliding max-prefill-size window size (#34856)
This commit is contained in:
+3
-1
@@ -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": "64",
|
"SGLANG_PREFILL_DELAYER_MAX_PREFILL_BS_WINDOW_SIZE": "128",
|
||||||
"PYTHONPATH": f"{MINIMAX_M2_5_EAGLE3_MODEL_PATH}:{os.environ.get('PYTHONPATH', '')}",
|
"PYTHONPATH": f"{MINIMAX_M2_5_EAGLE3_MODEL_PATH}:{os.environ.get('PYTHONPATH', '')}",
|
||||||
}
|
}
|
||||||
|
|
||||||
@@ -41,6 +41,8 @@ MINIMAX_M2_5_W8A8_4P_IN64K_OUT1K_PREFIX90_OTHER_ARGS = [
|
|||||||
"--tool-call-parser",
|
"--tool-call-parser",
|
||||||
"minimax-m2",
|
"minimax-m2",
|
||||||
"--enable-prefill-delayer",
|
"--enable-prefill-delayer",
|
||||||
|
"--prefill-delayer-max-delay-passes",
|
||||||
|
128,
|
||||||
"--prefill-max-requests",
|
"--prefill-max-requests",
|
||||||
10,
|
10,
|
||||||
"--chunked-prefill-size",
|
"--chunked-prefill-size",
|
||||||
|
|||||||
+1
@@ -20,6 +20,7 @@ QWEN3_8B_ENVS = {
|
|||||||
"SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "1",
|
"SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "1",
|
||||||
"SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE": "1",
|
"SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE": "1",
|
||||||
"SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES": "50",
|
"SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES": "50",
|
||||||
|
"SGLANG_PREFILL_DELAYER_MAX_PREFILL_BS_WINDOW_SIZE": "128",
|
||||||
}
|
}
|
||||||
|
|
||||||
QWEN3_8B_OTHER_ARGS = [
|
QWEN3_8B_OTHER_ARGS = [
|
||||||
|
|||||||
+2
-1
@@ -21,7 +21,8 @@ QWEN3_235B_ENVS = {
|
|||||||
"SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "1",
|
"SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "1",
|
||||||
"SGLANG_ENABLE_SPEC_V2": "1",
|
"SGLANG_ENABLE_SPEC_V2": "1",
|
||||||
"SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE": "1",
|
"SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE": "1",
|
||||||
"SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES": "100",
|
"SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES": "50",
|
||||||
|
"SGLANG_PREFILL_DELAYER_MAX_PREFILL_BS_WINDOW_SIZE": "16",
|
||||||
"SGLANG_NPU_PROFILING": "0",
|
"SGLANG_NPU_PROFILING": "0",
|
||||||
"SGLANG_NPU_PROFILING_BS": "27",
|
"SGLANG_NPU_PROFILING_BS": "27",
|
||||||
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "188416",
|
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "188416",
|
||||||
|
|||||||
+1
@@ -21,6 +21,7 @@ QWEN3_30B_A3B_ENVS = {
|
|||||||
"SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "1",
|
"SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "1",
|
||||||
"SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE": "1",
|
"SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE": "1",
|
||||||
"SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES": "200",
|
"SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES": "200",
|
||||||
|
"SGLANG_PREFILL_DELAYER_MAX_PREFILL_BS_WINDOW_SIZE": "128",
|
||||||
"DEEPEP_HCCL_BUFFSIZE": "400",
|
"DEEPEP_HCCL_BUFFSIZE": "400",
|
||||||
}
|
}
|
||||||
|
|
||||||
|
|||||||
+1
@@ -20,6 +20,7 @@ QWEN3_32B_ENVS = {
|
|||||||
"SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "1",
|
"SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "1",
|
||||||
"SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE": "1",
|
"SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE": "1",
|
||||||
"SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES": "100",
|
"SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES": "100",
|
||||||
|
"SGLANG_PREFILL_DELAYER_MAX_PREFILL_BS_WINDOW_SIZE": "128",
|
||||||
"SGLANG_NPU_USE_DEEPGEMM": "1",
|
"SGLANG_NPU_USE_DEEPGEMM": "1",
|
||||||
}
|
}
|
||||||
|
|
||||||
|
|||||||
+2
-1
@@ -22,7 +22,8 @@ QWEN3_6_27B_1024_ENVS = {
|
|||||||
"SGLANG_NPU_PROFILING_STAGE": "prefill",
|
"SGLANG_NPU_PROFILING_STAGE": "prefill",
|
||||||
"SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "1",
|
"SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "1",
|
||||||
"SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE": "1",
|
"SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE": "1",
|
||||||
"SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES": "150",
|
"SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES": "300",
|
||||||
|
"SGLANG_PREFILL_DELAYER_MAX_PREFILL_BS_WINDOW_SIZE": "128",
|
||||||
"ASCEND_USE_FIA": "1",
|
"ASCEND_USE_FIA": "1",
|
||||||
}
|
}
|
||||||
|
|
||||||
|
|||||||
+3
-1
@@ -25,7 +25,7 @@ QWEN3_6_27B_1080P_ENVS = {
|
|||||||
"SGLANG_NPU_PROFILING": "0",
|
"SGLANG_NPU_PROFILING": "0",
|
||||||
"SGLANG_NPU_PROFILING_STAGE": "prefill",
|
"SGLANG_NPU_PROFILING_STAGE": "prefill",
|
||||||
"SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "1",
|
"SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "1",
|
||||||
"SGLANG_PREFILL_DELAYER_MAX_PREFILL_BS_WINDOW_SIZE": "64",
|
"SGLANG_PREFILL_DELAYER_MAX_PREFILL_BS_WINDOW_SIZE": "128",
|
||||||
"ASCEND_USE_FIA": "1",
|
"ASCEND_USE_FIA": "1",
|
||||||
}
|
}
|
||||||
|
|
||||||
@@ -70,6 +70,8 @@ QWEN3_6_27B_1080P_OTHER_ARGS = [
|
|||||||
0.45,
|
0.45,
|
||||||
"--prefill-delayer-max-delay-ms",
|
"--prefill-delayer-max-delay-ms",
|
||||||
5500,
|
5500,
|
||||||
|
"--prefill-delayer-max-delay-passes",
|
||||||
|
300,
|
||||||
"--enable-multimodal",
|
"--enable-multimodal",
|
||||||
"--mm-attention-backend",
|
"--mm-attention-backend",
|
||||||
"ascend_attn",
|
"ascend_attn",
|
||||||
|
|||||||
Reference in New Issue
Block a user