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, 09:30 UTC
[Misc] Add chunked-prefill support on FlashInfer. (#9781)
Elfie Guo committed
Oct 30, 2024 at 09:33 UTC
9ff4511e43bb95efefd4e28048ca257e408277fb
2 files changed
+72
-28
tests/basic_correctness/test_chunked_prefill.py
+12
@@ -11,6 +11,8 @@ from contextlib import nullcontext
11
12
import pytest
13
14
+from tests.kernels.utils import override_backend_env_variable
15
+
16
from ..models.utils import check_logprobs_close, check_outputs_equal
17
from ..utils import multi_gpu_test
18
@@ -28,6 +30,7 @@ MODELS = [
30
# NOTE: Increasing this in this suite will fail CI because we currently cannot
31
# reset distributed env properly. Use a value > 1 just when you test.
32
@pytest.mark.parametrize("tensor_parallel_size", [1])
33
+@pytest.mark.parametrize("attention_backend", ["FLASHINFER", "FLASH_ATTN"])
34
def test_models(
35
hf_runner,
36
vllm_runner,
@@ -38,11 +41,15 @@ def test_models(
41
chunked_prefill_token_size: int,
42
enforce_eager: bool,
43
tensor_parallel_size: int,
44
+ attention_backend: str,
45
+ monkeypatch,
46
) -> None:
47
"""
48
Checks exact match decode between huggingface model and vllm runner with
49
chunked prefill.
50
"""
51
+ override_backend_env_variable(monkeypatch, attention_backend)
52
+
53
max_num_seqs = chunked_prefill_token_size
54
max_num_batched_tokens = chunked_prefill_token_size
55
@@ -71,13 +78,18 @@ def test_models(
78
@multi_gpu_test(num_gpus=2)
79
@pytest.mark.parametrize("distributed_executor_backend", ["ray", "mp"])
80
@pytest.mark.parametrize("model", MODELS)
81
+@pytest.mark.parametrize("attention_backend", ["FLASHINFER", "FLASH_ATTN"])
82
def test_models_distributed(
83
hf_runner,
84
vllm_runner,
85
example_prompts,
86
model: str,
87
distributed_executor_backend: str,
88
+ attention_backend: str,
89
+ monkeypatch,
90
) -> None:
91
+ override_backend_env_variable(monkeypatch, attention_backend)
92
+
93
if (model == "meta-llama/Llama-2-7b-hf"
94
and distributed_executor_backend == "ray"):
95
# test ray adag
vllm/attention/backends/flashinfer.py
+60
-28
@@ -268,6 +268,11 @@ class FlashInferMetadata(AttentionMetadata):
268
# Maximum sequence length among prefill batch. 0 if there are decoding
269
# requests only.
270
max_prefill_seq_len: int
271
+ # Number of query tokens for each request in the batch.
272
+ # Currently, we require that all requests have the same number of query
273
+ # tokens during the decoding phase. When speculavie decoding is enabled,
274
+ # decode_query_len might be greater than 1. In all other cases, it is 1.
275
+ decode_query_len: Optional[int] = 1
276
277
use_cuda_graph: bool = True
278
@@ -335,6 +340,7 @@ class FlashInferMetadata(AttentionMetadata):
340
assert self.paged_kv_last_page_len is not None
341
assert self.block_table_bound is not None
342
assert self.seq_lens_tensor is not None
343
+ self.query_start_loc = self.query_start_loc[:self.num_prefills + 1]
344
batch_size = self.query_start_loc.shape[0] - 1
345
assert batch_size >= 0
346
# We will use flash attention for profiling to
@@ -349,11 +355,13 @@ class FlashInferMetadata(AttentionMetadata):
355
self.paged_kv_indices = self.paged_kv_indices.to(self.device)
356
self.prefill_wrapper.end_forward()
357
self.prefill_wrapper.begin_forward(
352
- self.query_start_loc, self.paged_kv_indptr,
353
- self.paged_kv_indices, self.paged_kv_last_page_len,
358
+ self.query_start_loc,
359
+ self.paged_kv_indptr[:self.num_prefills + 1],
360
+ self.paged_kv_indices,
361
+ self.paged_kv_last_page_len[:self.num_prefills],
362
self.num_qo_heads, self.num_kv_heads, self.head_dim,
363
self.page_size)
356
- else:
364
+ if self.num_decode_tokens > 0:
365
assert self.paged_kv_indices is not None
366
assert self.paged_kv_indptr is not None
367
assert self.paged_kv_last_page_len is not None
@@ -370,9 +378,9 @@ class FlashInferMetadata(AttentionMetadata):
378
assert self.decode_wrapper is not None
379
self.decode_wrapper.end_forward()
380
self.decode_wrapper.begin_forward(
373
- self.paged_kv_indptr,
381
+ self.paged_kv_indptr[self.num_prefills:],
382
self.paged_kv_indices,
375
- self.paged_kv_last_page_len,
383
+ self.paged_kv_last_page_len[self.num_prefills:],
384
self.num_qo_heads,
385
self.num_kv_heads,
386
self.head_dim,
@@ -397,21 +405,14 @@ class FlashInferMetadata(AttentionMetadata):
405
406
@property
407
def prefill_metadata(self) -> Optional["FlashInferMetadata"]:
400
- # Currently chunked prefill is not supported
401
- if self.num_decode_tokens == 0:
402
- assert self.num_prefills > 0
403
- return self
404
-
405
- return None
408
+ if self.num_prefills == 0:
409
+ return None
410
+ return self
411
412
@property
413
def decode_metadata(self) -> Optional["FlashInferMetadata"]:
409
- # Currently chunked prefill is not supported
410
- if self.num_prefills > 0:
411
- assert self.num_decode_tokens == 0, (
412
- "Chunked prefill is not supported with flashinfer yet.")
414
+ if self.num_decode_tokens == 0:
415
return None
414
-
416
return self
417
418
def advance_step(self,
@@ -599,11 +600,12 @@ class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):
600
601
max_prefill_seq_len = max(self.prefill_seq_lens, default=0)
602
num_decode_tokens = self.num_decode_tokens
603
+ decode_query_len = max(query_lens[self.num_prefills:], default=1)
604
605
if use_captured_graph:
606
self.slot_mapping.extend([PAD_SLOT_ID] * cuda_graph_pad_size)
607
self.block_tables.extend([] * cuda_graph_pad_size)
606
- num_decode_tokens = batch_size
608
+ num_decode_tokens = batch_size - self.num_prefill_tokens
609
610
# The shape of graph_block_tables is
611
# [max batch size, max context len // block size].
@@ -689,6 +691,7 @@ class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):
691
self.runner.kv_cache_dtype, self.runner.model_config.dtype)
692
693
return FlashInferMetadata(
694
+ decode_query_len=decode_query_len,
695
num_prefills=self.num_prefills,
696
slot_mapping=slot_mapping_tensor,
697
num_prefill_tokens=self.num_prefill_tokens,
@@ -811,12 +814,6 @@ def unified_flash_infer(
814
key = key.view(-1, num_kv_heads, head_size)
815
value = value.view(-1, num_kv_heads, head_size)
816
814
- if attn_metadata.num_prefill_tokens > 0:
815
- assert attn_metadata.num_decode_tokens == 0, (
816
- "Chunked prefill is not supported with flashinfer yet.")
817
- if attn_metadata.num_decode_tokens > 0:
818
- assert attn_metadata.num_prefill_tokens == 0, (
819
- "Chunked prefill is not supported with flashinfer yet.")
817
if kv_cache.numel() > 0:
818
# Use the same reshape and cache kernel as flash attention.
819
ops.reshape_and_cache_flash(
@@ -836,14 +833,33 @@ def unified_flash_infer(
833
kv_cache_dtype)
834
kv_cache = kv_cache.view(torch_dtype)
835
836
+ num_prefill_tokens = attn_metadata.num_prefill_tokens
837
+ num_decode_tokens = attn_metadata.num_decode_tokens
838
+ assert key.shape[0] == num_prefill_tokens + num_decode_tokens, \
839
+ f"key : {key.shape} : #prefill tokens {num_prefill_tokens} : #decode tokens {num_decode_tokens}" # noqa
840
+ assert value.shape[0] == num_prefill_tokens + num_decode_tokens, \
841
+ f"value : {value.shape} : #prefill toks {num_prefill_tokens} : #decode toks {num_decode_tokens}" # noqa
842
query = query.contiguous() # Flashinfer requires query to be contiguous
843
+ # Query for decode. KV is not needed because it is already cached.
844
+ # QKV for prefill.
845
+ decode_query = query[num_prefill_tokens:]
846
+ query = query[:num_prefill_tokens]
847
+
848
+ key = key[:num_prefill_tokens]
849
+ value = value[:num_prefill_tokens]
850
+
851
+ assert query.shape[0] == num_prefill_tokens
852
+ assert decode_query.shape[0] == num_decode_tokens
853
+
854
+ prefill_output: Optional[torch.Tensor] = None
855
+ decode_output: Optional[torch.Tensor] = None
856
if prefill_meta := attn_metadata.prefill_metadata:
857
# We will use flash attention for prefill
858
# when kv_cache is not provided.
859
# This happens when vllm runs the profiling to
860
# determine the number of blocks.
861
if kv_cache.numel() == 0:
846
- output = flash_attn_varlen_func(
862
+ prefill_output = flash_attn_varlen_func(
863
q=query,
864
k=key,
865
v=value,
@@ -859,18 +875,34 @@ def unified_flash_infer(
875
else:
876
assert prefill_meta is not None
877
assert prefill_meta.prefill_wrapper is not None
862
- output = prefill_meta.prefill_wrapper.forward(
878
+ prefill_output = prefill_meta.prefill_wrapper.forward(
879
query, kv_cache, logits_soft_cap=logits_soft_cap, causal=True)
864
- else:
880
+ if decode_meta := attn_metadata.decode_metadata:
881
assert attn_metadata.decode_metadata is not None
882
assert attn_metadata.decode_metadata.decode_wrapper is not None
867
- output = attn_metadata.decode_metadata.decode_wrapper.forward(
868
- query,
883
+ decode_output = attn_metadata.decode_metadata.decode_wrapper.forward(
884
+ decode_query,
885
kv_cache,
886
sm_scale=softmax_scale,
887
logits_soft_cap=logits_soft_cap,
888
k_scale=k_scale,
889
v_scale=v_scale)
890
+
891
+ if prefill_output is None and decode_output is not None:
892
+ # Decode only batch.
893
+ output, num_tokens = decode_output, num_decode_tokens
894
+ elif decode_output is None and prefill_output is not None:
895
+ # Prefill only batch.
896
+ output, num_tokens = prefill_output, num_prefill_tokens
897
+ else:
898
+ # Chunked prefill batch does not work with speculative decoding in
899
+ # FlashInfer backend, so the query length for decode should be 1.
900
+ assert prefill_output is not None
901
+ assert decode_output is not None
902
+ assert decode_meta is not None
903
+ assert decode_meta.decode_query_len == 1
904
+ decode_output = decode_output.squeeze(1)
905
+ output = torch.cat([prefill_output, decode_output], dim=0)
906
return output.view(num_tokens, hidden_size)
907
908