Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[Bugfix] fix flashinfer cudagraph capture for PP #6708

Merged
merged 3 commits into from
Jul 24, 2024
Merged

[Bugfix] fix flashinfer cudagraph capture for PP #6708

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
5cfb57a
fix flashinfer cudagraph capture for PP
SolitaryThinker Jul 23, 2024
cb419ee
add test for PP+cudagraph with different attn backends
SolitaryThinker Jul 23, 2024
3a5c34e
format
SolitaryThinker Jul 24, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
24 changes: 24 additions & 0 deletions tests/distributed/test_pipeline_parallel.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -61,3 +61,27 @@ def test_compare_tp(TP_SIZE, PP_SIZE, EAGER_MODE, CHUNKED_PREFILL, MODEL_NAME,
tp_args.append("--enforce-eager")

compare_two_settings(MODEL_NAME, pp_args, tp_args)


@pytest.mark.parametrize("PP_SIZE, MODEL_NAME", [
(2, "JackFram/llama-160m"),
])
@pytest.mark.parametrize("ATTN_BACKEND", [
"FLASH_ATTN",
"FLASHINFER",
])
def test_pp_cudagraph(PP_SIZE, MODEL_NAME, ATTN_BACKEND):
cudagraph_args = [
# use half precision for speed and memory savings in CI environment
"--dtype",
"float16",
"--pipeline-parallel-size",
str(PP_SIZE),
"--distributed-executor-backend",
"ray",
]
os.environ["VLLM_ATTENTION_BACKEND"] = ATTN_BACKEND

eager_args = cudagraph_args + ["--enforce-eager"]

compare_two_settings(MODEL_NAME, eager_args, cudagraph_args)
14 changes: 7 additions & 7 deletions vllm/worker/model_runner.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1040,9 +1040,9 @@ def capture_model(self, kv_caches: List[List[torch.Tensor]]) -> None:
self.parallel_config.pipeline_parallel_size):
for batch_size in reversed(batch_size_capture_list):
if self.attn_backend.get_name() == "flashinfer":
indptr_buffer = indptr_buffer[:batch_size + 1]
last_page_len_buffer = last_page_len_buffer[:
batch_size]
_indptr_buffer = indptr_buffer[:batch_size + 1]
_last_page_len_buffer = last_page_len_buffer[:
batch_size]

num_qo_heads = (
self.model_config.get_num_attention_heads(
Expand All @@ -1055,8 +1055,8 @@ def capture_model(self, kv_caches: List[List[torch.Tensor]]) -> None:
use_tensor_cores = False
decode_wrapper = \
CUDAGraphBatchDecodeWithPagedKVCacheWrapper(
decode_workspace_buffer, indptr_buffer,
indices_buffer, last_page_len_buffer, "NHD",
decode_workspace_buffer, _indptr_buffer,
indices_buffer, _last_page_len_buffer, "NHD",
use_tensor_cores)
kv_cache_dtype = get_kv_cache_torch_dtype(
self.kv_cache_dtype, self.model_config.dtype)
Expand Down Expand Up @@ -1131,10 +1131,10 @@ def capture_model(self, kv_caches: List[List[torch.Tensor]]) -> None:
self.model, self.attn_backend.get_name())

if self.attn_backend.get_name() == "flashinfer":
graph_runner.flashinfer_indptr_buffer = indptr_buffer
graph_runner.flashinfer_indptr_buffer = _indptr_buffer
graph_runner.flashinfer_indices_buffer = indices_buffer
graph_runner.flashinfer_last_page_len_buffer = \
last_page_len_buffer
_last_page_len_buffer
graph_runner.flashinfer_decode_workspace_buffer = \
decode_workspace_buffer
graph_runner.flashinfer_decode_wrapper = \
Expand Down
Loading