A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 16, 20:45 UTC
[Bugfix] fix flashinfer cudagraph capture for PP (#6708)
William Lin committed
Jul 23, 2024 at 18:49 UTC
5e8ca973ebd5584582923b8ed1d3d823769a80a5
2 files changed
+31
-7
tests/distributed/test_pipeline_parallel.py
+24
@@ -61,3 +61,27 @@ def test_compare_tp(TP_SIZE, PP_SIZE, EAGER_MODE, CHUNKED_PREFILL, MODEL_NAME,
61
tp_args.append("--enforce-eager")
62
63
compare_two_settings(MODEL_NAME, pp_args, tp_args)
64
+
65
+
66
+@pytest.mark.parametrize("PP_SIZE, MODEL_NAME", [
67
+ (2, "JackFram/llama-160m"),
68
+])
69
+@pytest.mark.parametrize("ATTN_BACKEND", [
70
+ "FLASH_ATTN",
71
+ "FLASHINFER",
72
+])
73
+def test_pp_cudagraph(PP_SIZE, MODEL_NAME, ATTN_BACKEND):
74
+ cudagraph_args = [
75
+ # use half precision for speed and memory savings in CI environment
76
+ "--dtype",
77
+ "float16",
78
+ "--pipeline-parallel-size",
79
+ str(PP_SIZE),
80
+ "--distributed-executor-backend",
81
+ "ray",
82
+ ]
83
+ os.environ["VLLM_ATTENTION_BACKEND"] = ATTN_BACKEND
84
+
85
+ eager_args = cudagraph_args + ["--enforce-eager"]
86
+
87
+ compare_two_settings(MODEL_NAME, eager_args, cudagraph_args)
vllm/worker/model_runner.py
+7
-7
@@ -1040,9 +1040,9 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
1040
self.parallel_config.pipeline_parallel_size):
1041
for batch_size in reversed(batch_size_capture_list):
1042
if self.attn_backend.get_name() == "flashinfer":
1043
- indptr_buffer = indptr_buffer[:batch_size + 1]
1044
- last_page_len_buffer = last_page_len_buffer[:
1045
- batch_size]
1043
+ _indptr_buffer = indptr_buffer[:batch_size + 1]
1044
+ _last_page_len_buffer = last_page_len_buffer[:
1045
+ batch_size]
1046
1047
num_qo_heads = (
1048
self.model_config.get_num_attention_heads(
@@ -1055,8 +1055,8 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
1055
use_tensor_cores = False
1056
decode_wrapper = \
1057
CUDAGraphBatchDecodeWithPagedKVCacheWrapper(
1058
- decode_workspace_buffer, indptr_buffer,
1059
- indices_buffer, last_page_len_buffer, "NHD",
1058
+ decode_workspace_buffer, _indptr_buffer,
1059
+ indices_buffer, _last_page_len_buffer, "NHD",
1060
use_tensor_cores)
1061
kv_cache_dtype = get_kv_cache_torch_dtype(
1062
self.kv_cache_dtype, self.model_config.dtype)
@@ -1131,10 +1131,10 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
1131
self.model, self.attn_backend.get_name())
1132
1133
if self.attn_backend.get_name() == "flashinfer":
1134
- graph_runner.flashinfer_indptr_buffer = indptr_buffer
1134
+ graph_runner.flashinfer_indptr_buffer = _indptr_buffer
1135
graph_runner.flashinfer_indices_buffer = indices_buffer
1136
graph_runner.flashinfer_last_page_len_buffer = \
1137
- last_page_len_buffer
1137
+ _last_page_len_buffer
1138
graph_runner.flashinfer_decode_workspace_buffer = \
1139
decode_workspace_buffer
1140
graph_runner.flashinfer_decode_wrapper = \