[Bugfix][310p] the new A5 mmencoder op donot support 310p (#7518)
### What this PR does / why we need it?
Because the new A5 MMEncoder operator was merged, the 310P can no longer
run any VL models. This PR fixes that issue. details at #7046
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
e2e
- vLLM version: v0.17.0
- vLLM main:
8b6325758c
---------
Signed-off-by: Tflowers-0129 <2906339855@qq.com>
This commit is contained in:
81
tests/ut/_310p/ops/test_mm_encoder_attention_310.py
Normal file
81
tests/ut/_310p/ops/test_mm_encoder_attention_310.py
Normal file
@@ -0,0 +1,81 @@
|
|||||||
|
#
|
||||||
|
# Copyright (c) 2026 Huawei Technologies Co., Ltd. All Rights Reserved.
|
||||||
|
#
|
||||||
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||||
|
# you may not use this file except in compliance with the License.
|
||||||
|
# You may obtain a copy of the License at
|
||||||
|
#
|
||||||
|
# http://www.apache.org/licenses/LICENSE-2.0
|
||||||
|
#
|
||||||
|
# Unless required by applicable law or agreed to in writing, software
|
||||||
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||||
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||||
|
# See the License for the specific language governing permissions and
|
||||||
|
# limitations under the License.
|
||||||
|
|
||||||
|
from unittest import mock
|
||||||
|
|
||||||
|
import torch
|
||||||
|
|
||||||
|
from vllm_ascend import utils
|
||||||
|
from vllm_ascend._310p.ops.mm_encoder_attention import AscendMMEncoderAttention310
|
||||||
|
|
||||||
|
|
||||||
|
def test_register_customop_overrides_mm_encoder_attention_for_310p():
|
||||||
|
original_registered = utils._ASCEND_CUSTOMOP_IS_REIGISTERED
|
||||||
|
try:
|
||||||
|
utils._ASCEND_CUSTOMOP_IS_REIGISTERED = False
|
||||||
|
with (
|
||||||
|
mock.patch("vllm.model_executor.custom_op.CustomOp.register_oot"),
|
||||||
|
mock.patch("vllm_ascend.utils.is_310p", return_value=True),
|
||||||
|
):
|
||||||
|
utils.register_ascend_customop()
|
||||||
|
|
||||||
|
assert utils.REGISTERED_ASCEND_OPS["MMEncoderAttention"] is AscendMMEncoderAttention310
|
||||||
|
finally:
|
||||||
|
utils._ASCEND_CUSTOMOP_IS_REIGISTERED = original_registered
|
||||||
|
|
||||||
|
|
||||||
|
def test_mm_encoder_attention_310_forward_oot_with_padding():
|
||||||
|
layer = AscendMMEncoderAttention310.__new__(AscendMMEncoderAttention310)
|
||||||
|
layer.num_heads = 4
|
||||||
|
layer.num_kv_heads = 2
|
||||||
|
layer.head_size = 80
|
||||||
|
layer.enable_pad = True
|
||||||
|
layer.scale_value = layer.head_size**-0.5
|
||||||
|
|
||||||
|
bsz, q_len, kv_len = 2, 3, 3
|
||||||
|
query = torch.randn(bsz, q_len, layer.num_heads, layer.head_size)
|
||||||
|
key = torch.randn(bsz, kv_len, layer.num_kv_heads, layer.head_size)
|
||||||
|
value = torch.randn(bsz, kv_len, layer.num_kv_heads, layer.head_size)
|
||||||
|
|
||||||
|
capture = {}
|
||||||
|
|
||||||
|
def fake_flash_attention_unpad(*, query, key, value, seq_len, scale_value, num_heads, num_kv_heads, out):
|
||||||
|
capture["query_shape"] = query.shape
|
||||||
|
capture["key_shape"] = key.shape
|
||||||
|
capture["value_shape"] = value.shape
|
||||||
|
capture["seq_len"] = seq_len
|
||||||
|
capture["scale_value"] = scale_value
|
||||||
|
capture["num_heads"] = num_heads
|
||||||
|
capture["num_kv_heads"] = num_kv_heads
|
||||||
|
out.copy_(query + 1.0)
|
||||||
|
|
||||||
|
with mock.patch(
|
||||||
|
"vllm_ascend._310p.ops.mm_encoder_attention.torch_npu._npu_flash_attention_unpad",
|
||||||
|
side_effect=fake_flash_attention_unpad,
|
||||||
|
create=True,
|
||||||
|
):
|
||||||
|
out = layer.forward_oot(query, key, value)
|
||||||
|
|
||||||
|
assert capture["query_shape"] == (bsz * q_len, layer.num_heads, 128)
|
||||||
|
assert capture["key_shape"] == (bsz * kv_len, layer.num_heads, 128)
|
||||||
|
assert capture["value_shape"] == (bsz * kv_len, layer.num_heads, 128)
|
||||||
|
assert capture["seq_len"].device.type == "cpu"
|
||||||
|
torch.testing.assert_close(capture["seq_len"], torch.tensor([q_len, q_len], dtype=torch.int32))
|
||||||
|
assert capture["num_heads"] == layer.num_heads
|
||||||
|
assert capture["num_kv_heads"] == layer.num_kv_heads
|
||||||
|
|
||||||
|
assert out.shape == query.shape
|
||||||
|
torch.testing.assert_close(out, query + 1.0)
|
||||||
|
|
||||||
142
vllm_ascend/_310p/ops/mm_encoder_attention.py
Normal file
142
vllm_ascend/_310p/ops/mm_encoder_attention.py
Normal file
@@ -0,0 +1,142 @@
|
|||||||
|
#
|
||||||
|
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
|
||||||
|
# This file is a part of the vllm-ascend project.
|
||||||
|
#
|
||||||
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||||
|
# you may not use this file except in compliance with the License.
|
||||||
|
# You may obtain a copy of the License at
|
||||||
|
#
|
||||||
|
# http://www.apache.org/licenses/LICENSE-2.0
|
||||||
|
#
|
||||||
|
# Unless required by applicable law or agreed to in writing, software
|
||||||
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||||
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||||
|
# See the License for the specific language governing permissions and
|
||||||
|
# limitations under the License.
|
||||||
|
#
|
||||||
|
|
||||||
|
import einops
|
||||||
|
import torch
|
||||||
|
import torch.nn.functional as F
|
||||||
|
import torch_npu
|
||||||
|
from vllm.model_executor.layers.attention.mm_encoder_attention import MMEncoderAttention # type: ignore
|
||||||
|
|
||||||
|
MIN_PAD_SIZE: int = 64 # min_size to pad weight
|
||||||
|
MAX_PAD_SIZE: int = 128 # max_size to pad weight
|
||||||
|
|
||||||
|
# Use seq_lens CPU cache to avoid frequent d2h copy.
|
||||||
|
# AscendMMEncoderAttention310 will copy the cu_seqlens from NPU to CPU in every
|
||||||
|
# forward, since the op _npu_flash_attention_unpad() requires CPU cu_seqlens
|
||||||
|
# (otherwise it will break down).
|
||||||
|
# Thus, we use seq_lens_cpu_cache to cache this tensor, since it's shared
|
||||||
|
# between all layers, but may change in different forward step. When the
|
||||||
|
# current layer_index is 0, we update the cache, otherwise we directly use the
|
||||||
|
# cache to avoid frequent diff and copy operations, which are costful.
|
||||||
|
seq_lens_cpu_cache: torch.Tensor = None
|
||||||
|
|
||||||
|
|
||||||
|
class AscendMMEncoderAttention310(MMEncoderAttention):
|
||||||
|
def __init__(
|
||||||
|
self,
|
||||||
|
num_heads: int,
|
||||||
|
head_size: int,
|
||||||
|
scale: float | None = None,
|
||||||
|
num_kv_heads: int | None = None,
|
||||||
|
prefix: str = "",
|
||||||
|
) -> None:
|
||||||
|
"""
|
||||||
|
Args:
|
||||||
|
num_heads: number of attention heads per partition.
|
||||||
|
head_size: hidden_size per attention head.
|
||||||
|
scale: scale factor.
|
||||||
|
num_kv_heads: number of kv heads.
|
||||||
|
prefix: This has no effect, it is only here to make it easier to
|
||||||
|
swap between Attention and MMEncoderAttention.
|
||||||
|
multimodal_config: configs for multi-modal.
|
||||||
|
"""
|
||||||
|
super().__init__(
|
||||||
|
num_heads=num_heads,
|
||||||
|
head_size=head_size,
|
||||||
|
scale=scale,
|
||||||
|
num_kv_heads=num_kv_heads,
|
||||||
|
prefix=prefix,
|
||||||
|
)
|
||||||
|
|
||||||
|
self.enable_pad = self.head_size > MIN_PAD_SIZE and self.head_size < MAX_PAD_SIZE
|
||||||
|
self.scale_value = self.head_size**-0.5
|
||||||
|
|
||||||
|
def _reshape_qkv_to_3d(
|
||||||
|
self,
|
||||||
|
query: torch.Tensor,
|
||||||
|
key: torch.Tensor,
|
||||||
|
value: torch.Tensor,
|
||||||
|
bsz: int,
|
||||||
|
q_len: int,
|
||||||
|
kv_len: int,
|
||||||
|
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
|
||||||
|
"""
|
||||||
|
Reshape query, key, value to 3D tensors:
|
||||||
|
(batch_size * seq_len, num_heads, head_size)
|
||||||
|
"""
|
||||||
|
query = query.view(bsz * q_len, self.num_heads, self.head_size)
|
||||||
|
key = key.view(bsz * kv_len, self.num_kv_heads, self.head_size)
|
||||||
|
value = value.view(bsz * kv_len, self.num_kv_heads, self.head_size)
|
||||||
|
self.num_queries_per_kv = self.num_heads // self.num_kv_heads
|
||||||
|
if (num_repeat := self.num_queries_per_kv) > 1:
|
||||||
|
# Handle MQA and GQA
|
||||||
|
key = torch.repeat_interleave(key, num_repeat, dim=1)
|
||||||
|
value = torch.repeat_interleave(value, num_repeat, dim=1)
|
||||||
|
|
||||||
|
return query, key, value
|
||||||
|
|
||||||
|
def forward_oot(
|
||||||
|
self,
|
||||||
|
query: torch.Tensor,
|
||||||
|
key: torch.Tensor,
|
||||||
|
value: torch.Tensor,
|
||||||
|
cu_seqlens: torch.Tensor | None = None,
|
||||||
|
max_seqlen: torch.Tensor | None = None, # Only used for Flash Attention
|
||||||
|
sequence_lengths: torch.Tensor | None = None,
|
||||||
|
):
|
||||||
|
bsz, q_len = query.size()[:2]
|
||||||
|
kv_len = key.size(1)
|
||||||
|
is_reshaped = query.dim() == 4
|
||||||
|
|
||||||
|
# Directly use seq_lens cpu cache to avoid d2h copy.
|
||||||
|
if cu_seqlens is None:
|
||||||
|
cu_seqlens = torch.arange(0, (bsz + 1) * q_len, step=q_len, dtype=torch.int32, device="cpu")
|
||||||
|
seq_lens_cpu = torch.diff(cu_seqlens).to("cpu")
|
||||||
|
|
||||||
|
# q, k, v: [b, s, head, head_dim] -> [b * s, head, head_dim]
|
||||||
|
q, k, v = self._reshape_qkv_to_3d(query, key, value, bsz, q_len, kv_len)
|
||||||
|
|
||||||
|
if self.enable_pad:
|
||||||
|
origin_shape = q.shape[-1]
|
||||||
|
pad_len = MAX_PAD_SIZE - origin_shape
|
||||||
|
# [b * s, head, head_dim] -> [b * s, head, MAX_PAD_SIZE]
|
||||||
|
q = F.pad(q, (0, pad_len), mode="constant", value=0)
|
||||||
|
k = F.pad(k, (0, pad_len), mode="constant", value=0)
|
||||||
|
v = F.pad(v, (0, pad_len), mode="constant", value=0)
|
||||||
|
|
||||||
|
context_layer = torch.empty_like(q)
|
||||||
|
|
||||||
|
# operator requires pta version >= 2.5.1
|
||||||
|
torch_npu._npu_flash_attention_unpad(
|
||||||
|
query=q,
|
||||||
|
key=k,
|
||||||
|
value=v,
|
||||||
|
seq_len=seq_lens_cpu,
|
||||||
|
scale_value=self.scale_value,
|
||||||
|
num_heads=self.num_heads,
|
||||||
|
num_kv_heads=self.num_kv_heads,
|
||||||
|
out=context_layer,
|
||||||
|
)
|
||||||
|
|
||||||
|
if self.enable_pad:
|
||||||
|
context_layer = context_layer[..., :origin_shape]
|
||||||
|
|
||||||
|
if is_reshaped:
|
||||||
|
context_layer = einops.rearrange(context_layer, "(b s) h d -> b s h d", b=bsz).contiguous()
|
||||||
|
else:
|
||||||
|
context_layer = einops.rearrange(context_layer, "(b s) h d -> b s (h d)", b=bsz).contiguous()
|
||||||
|
return context_layer
|
||||||
@@ -662,6 +662,7 @@ def register_ascend_customop(vllm_config: VllmConfig | None = None):
|
|||||||
from vllm_ascend._310p.fused_moe.fused_moe import AscendFusedMoE310, AscendSharedFusedMoE310
|
from vllm_ascend._310p.fused_moe.fused_moe import AscendFusedMoE310, AscendSharedFusedMoE310
|
||||||
from vllm_ascend._310p.ops.activation import AscendSiluAndMul310
|
from vllm_ascend._310p.ops.activation import AscendSiluAndMul310
|
||||||
from vllm_ascend._310p.ops.layernorm import AscendGemmaRMSNorm310, AscendRMSNorm310
|
from vllm_ascend._310p.ops.layernorm import AscendGemmaRMSNorm310, AscendRMSNorm310
|
||||||
|
from vllm_ascend._310p.ops.mm_encoder_attention import AscendMMEncoderAttention310
|
||||||
from vllm_ascend._310p.ops.rotary_embedding import AscendRotaryEmbedding310
|
from vllm_ascend._310p.ops.rotary_embedding import AscendRotaryEmbedding310
|
||||||
from vllm_ascend._310p.ops.vocab_parallel_embedding import (
|
from vllm_ascend._310p.ops.vocab_parallel_embedding import (
|
||||||
AscendParallelLMHead310,
|
AscendParallelLMHead310,
|
||||||
@@ -678,6 +679,7 @@ def register_ascend_customop(vllm_config: VllmConfig | None = None):
|
|||||||
"SharedFusedMoE": AscendSharedFusedMoE310,
|
"SharedFusedMoE": AscendSharedFusedMoE310,
|
||||||
"ParallelLMHead": AscendParallelLMHead310,
|
"ParallelLMHead": AscendParallelLMHead310,
|
||||||
"VocabParallelEmbedding": AscendVocabParallelEmbedding310,
|
"VocabParallelEmbedding": AscendVocabParallelEmbedding310,
|
||||||
|
"MMEncoderAttention": AscendMMEncoderAttention310,
|
||||||
}
|
}
|
||||||
)
|
)
|
||||||
|
|
||||||
|
|||||||
Reference in New Issue
Block a user