biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[FlashInfer] Upgrade to 0.2.0 (#11194)

Browse Source 
Signed-off-by: Bowen Wang <abmfy@icloud.com>
Signed-off-by: youkaichao <youkaichao@gmail.com>
Co-authored-by: youkaichao <youkaichao@gmail.com>
This commit is contained in:
Bowen Wang
2025-01-28 02:19:24 +08:00
committed by GitHub
parent 3f1fc7425a
commit 2bc3fbba0c
10 changed files with 256 additions and 74 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
tests/basic_correctness/test_basic_correctness.py
View File

@@ -61,9 +61,10 @@ def test_models(
if backend == "FLASHINFER" and current_platform.is_rocm():
pytest.skip("Flashinfer does not support ROCm/HIP.")
if backend == "XFORMERS" and model == "google/gemma-2-2b-it":
if backend in ("XFORMERS",
"FLASHINFER") and model == "google/gemma-2-2b-it":
pytest.skip(
"XFORMERS does not support gemma2 with full context length.")
f"{backend} does not support gemma2 with full context length.")
os.environ["VLLM_ATTENTION_BACKEND"] = backend

2
tests/compile/test_basic_correctness.py
View File

@@ -58,7 +58,7 @@ test_settings = [
model_args=["--task", "embed"],
pp_size=1,
tp_size=1,
attn_backend="FLASHINFER",
attn_backend="FLASH_ATTN",
method="encode",
fullgraph=True,
),

72
tests/kernels/test_flashinfer.py
View File

@@ -133,17 +133,19 @@ def test_flashinfer_decode_with_paged_kv(
use_tensor_cores=(
(num_query_heads//num_kv_heads) > 4)
)
wrapper.begin_forward(kv_indptr,
kv_indices,
kv_last_page_lens,
num_query_heads,
num_kv_heads,
head_size,
block_size,
"NONE",
data_type=dtype)
wrapper.plan(kv_indptr,
kv_indices,
kv_last_page_lens,
num_query_heads,
num_kv_heads,
head_size,
block_size,
"NONE",
q_data_type=dtype,
kv_data_type=dtype,
logits_soft_cap=soft_cap)
output = wrapper.forward(query, key_value_cache, logits_soft_cap=soft_cap)
output = wrapper.run(query, key_value_cache)
ref_output = ref_paged_attn(query=query,
key_cache=key_cache,
@@ -228,7 +230,7 @@ def test_flashinfer_prefill_with_paged_kv(seq_lens: List[Tuple[int, int]],
workspace_buffer = torch.empty(128 * 1024 * 1024, dtype=torch.int8)
wrapper = flashinfer.BatchPrefillWithPagedKVCacheWrapper(
workspace_buffer, "NHD")
wrapper.begin_forward(
wrapper.plan(
qo_indptr,
kv_indptr,
kv_indices,
@@ -237,12 +239,14 @@ def test_flashinfer_prefill_with_paged_kv(seq_lens: List[Tuple[int, int]],
num_kv_heads,
head_size,
block_size,
q_data_type=dtype,
kv_data_type=dtype,
logits_soft_cap=soft_cap,
)
output = wrapper.forward(
output = wrapper.run(
query,
key_value_cache,
logits_soft_cap=soft_cap,
)
ref_output = ref_paged_attn(query=query,
@@ -253,7 +257,7 @@ def test_flashinfer_prefill_with_paged_kv(seq_lens: List[Tuple[int, int]],
block_tables=block_tables,
scale=scale,
soft_cap=soft_cap)
torch.testing.assert_close(output, ref_output, atol=1e-2, rtol=1e-2), \
torch.testing.assert_close(output, ref_output, atol=5e-2, rtol=1e-2), \
f"{torch.max(torch.abs(output - ref_output))}"
@@ -332,7 +336,7 @@ def test_flashinfer_prefill_with_paged_fp8_kv(
workspace_buffer = torch.empty(128 * 1024 * 1024, dtype=torch.int8)
wrapper = flashinfer.BatchPrefillWithPagedKVCacheWrapper(
workspace_buffer, "NHD")
wrapper.begin_forward(
wrapper.plan(
qo_indptr,
kv_indptr,
kv_indices,
@@ -341,13 +345,12 @@ def test_flashinfer_prefill_with_paged_fp8_kv(
num_kv_heads,
head_size,
block_size,
q_data_type=dtype,
kv_data_type=kv_cache_dtype,
logits_soft_cap=soft_cap,
)
output = wrapper.forward(query,
kv_cache_fp8,
logits_soft_cap=soft_cap,
k_scale=k_scale,
v_scale=v_scale)
output = wrapper.run(query, kv_cache_fp8, k_scale=k_scale, v_scale=v_scale)
ref_output = ref_paged_attn(query=query,
key_cache=key_cache.squeeze(1),
@@ -360,7 +363,7 @@ def test_flashinfer_prefill_with_paged_fp8_kv(
del query
del block_tables
# verify prefill fp8
torch.testing.assert_close(output, ref_output, atol=1e-2, rtol=1e-2), \
torch.testing.assert_close(output, ref_output, atol=5e-2, rtol=1e-2), \
f"{torch.max(torch.abs(output - ref_output))}"
@@ -439,21 +442,18 @@ def test_flashinfer_decode_with_paged_fp8_kv(
wrapper = flashinfer.\
BatchDecodeWithPagedKVCacheWrapper(workspace_buffer, "NHD",
use_tensor_cores=use_tensor_cores)
wrapper.begin_forward(kv_indptr,
kv_indices,
kv_last_page_lens,
num_query_heads,
num_kv_heads,
head_size,
block_size,
"NONE",
data_type=dtype,
q_data_type=dtype)
output = wrapper.forward(query,
kv_cache_fp8,
logits_soft_cap=soft_cap,
k_scale=k_scale,
v_scale=v_scale)
wrapper.plan(kv_indptr,
kv_indices,
kv_last_page_lens,
num_query_heads,
num_kv_heads,
head_size,
block_size,
"NONE",
q_data_type=dtype,
kv_data_type=kv_cache_dtype,
logits_soft_cap=soft_cap)
output = wrapper.run(query, kv_cache_fp8, k_scale=k_scale, v_scale=v_scale)
key_cache = key_value_cache[:, 0, :, :, :].squeeze(1)
value_cache = key_value_cache[:, 1, :, :, :].squeeze(1)