Files
xc-llm-ascend/vllm_ascend/_310p/model_runner_310p.py
linfeng-yuan 68d8d20ca2 [misc] move mxfp_compat into device to decouple from quantization init chain (#6918)
### What this PR does / why we need it?
`mxfp_compat` only provides dtype/symbol compatibility helpers for
different `torch_npu` versions, but it was placed under
`vllm_ascend.quantization`. Importing it from device/ops paths could
trigger `quantization/__init__.py` and pull in heavy quantization method
dependencies, increasing startup coupling and causing import-cycle risk
(especially on 310P paths).

### Does this PR introduce _any_ user-facing change?
No functional behavior change intended.

### How was this patch tested?
CI passed.

- vLLM version: v0.16.0
- vLLM main:
15d76f74e2

---------

Signed-off-by: linfeng-yuan <1102311262@qq.com>
2026-03-02 18:17:01 +08:00

284 lines
14 KiB
Python

#
# Copyright (c) 2025 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.
# This file is a part of the vllm-ascend project.
#
from __future__ import annotations
import numpy as np
import torch
import torch_npu
from vllm.logger import logger
from vllm.v1.core.sched.output import SchedulerOutput
from vllm.v1.kv_cache_interface import AttentionSpec, KVCacheConfig, MambaSpec
from vllm_ascend.utils import ACL_FORMAT_FRACTAL_NZ
from vllm_ascend.worker.model_runner_v1 import NPUModelRunner
class NPUModelRunner310(NPUModelRunner):
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self._acl_format = ACL_FORMAT_FRACTAL_NZ
def initialize_kv_cache_tensors(self, kv_cache_config: KVCacheConfig) -> dict[str, torch.Tensor]:
"""
Initialize the memory buffer for KV cache.
Args:
kv_cache_config: The KV cache config
Returns:
Dict[str, torch.Tensor]: A map between layer names to their
corresponding memory buffer for KV cache.
"""
# 310P limitation: KV transfer is not supported
if self.vllm_config.kv_transfer_config is not None:
raise ValueError("KV cache transfer is not supported for 310P.")
if self.use_sparse:
raise ValueError("Deepseek Sparse Attention is not supported for 310P.")
if self.model_config.use_mla:
raise ValueError("MLAAttention is not supported for 310P.")
# Initialize the memory size for KV cache
kv_cache_size = self._calculate_kv_cache_tensors_size(kv_cache_config)
# Allocate and reshape KV cache Tensors
kv_caches = self._allocate_kv_cache_and_reshape_tensors(kv_cache_config, kv_cache_size)
# Set up cross-layer KV cache sharing
for layer_name, target_layer_name in self.shared_kv_cache_layers.items():
logger.debug("%s reuses KV cache of %s", layer_name, target_layer_name)
kv_caches[layer_name] = kv_caches[target_layer_name]
from vllm.v1.worker.utils import bind_kv_cache
bind_kv_cache(kv_caches, self.compilation_config.static_forward_context, self.kv_caches)
return kv_caches
def _calculate_kv_cache_tensors_size(self, kv_cache_config: KVCacheConfig) -> dict[str, int]:
"""
Initializes the KV cache size. The buffer needs to be reshaped to the desired shape before being used by
the models.
Args:
kv_cache_config: The KV cache config
Returns:
dict[str, int]: A map between layer names to their
corresponding memory buffer size.
"""
# init kv cache tensors
kv_cache_sizes: dict[str, int] = {}
for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
# TODO: REFACTOR ME to sharing hybrid cache
for idx in range(len(kv_cache_tensor.shared_by)):
layer_name = kv_cache_tensor.shared_by[idx]
if "linear_attn" in layer_name and layer_name not in kv_cache_sizes:
# for mamba linear attention
kv_cache_size = kv_cache_tensor.size
for layer_name_inner in kv_cache_tensor.shared_by:
# shared the kvcache between the self_attn specs in the same group
if "linear_attn" in layer_name_inner:
kv_cache_sizes[layer_name_inner] = kv_cache_size
elif "attn" in layer_name and layer_name not in kv_cache_sizes:
kv_tensor_split_factor = 2
kv_tensor_size = int(kv_cache_tensor.size // kv_tensor_split_factor)
for layer_name_inner in kv_cache_tensor.shared_by:
# shared the kvcache between the self_attn specs in the same group
if "attn" in layer_name_inner and "linear_attn" not in layer_name_inner:
kv_cache_sizes[layer_name_inner] = kv_tensor_size
layer_names = set()
for group in kv_cache_config.kv_cache_groups:
for layer_name in group.layer_names:
if layer_name in self.runner_only_attn_layers:
continue
layer_names.add(layer_name)
assert layer_names == set(kv_cache_sizes.keys()), "Some layers are not correctly initialized"
return kv_cache_sizes
def _allocate_kv_cache_and_reshape_tensors(
self,
kv_cache_config: KVCacheConfig,
kv_cache_sizes: dict[str, int],
) -> dict[str, torch.Tensor]:
"""
Allocate the KV cache tensors to the desired shape and dtype.
Args:
kv_cache_config: The KV cache config
kv_cache_sizes: The KV cache size of each layer
Returns:
dict[str, torch.Tensor]: A map between layer names to their
corresponding memory buffer for KV cache.
"""
kv_caches: dict[str, torch.Tensor] = {}
for group in self._kv_cache_spec_attn_group_iterator():
kv_cache_spec = group.kv_cache_spec
attn_backend = group.backend
for layer_name in group.layer_names:
if layer_name in self.runner_only_attn_layers:
continue
if isinstance(kv_cache_spec, AttentionSpec):
kv_tensor_size = kv_cache_sizes[layer_name]
assert kv_tensor_size is not None
sum_page_size_bytes = kv_tensor_size * 2
assert sum_page_size_bytes % kv_cache_spec.page_size_bytes == 0
num_blocks = sum_page_size_bytes // kv_cache_spec.page_size_bytes
assert num_blocks >= kv_cache_config.num_blocks
if hasattr(attn_backend, "get_supported_kernel_block_sizes") and self.use_hybrid_blocks:
block_size = attn_backend.get_supported_kernel_block_sizes()[0]
block_size_chunk = kv_cache_spec.block_size // block_size
kv_cache_shape = attn_backend.get_kv_cache_shape(
num_blocks * block_size_chunk,
block_size,
kv_cache_spec.num_kv_heads,
kv_cache_spec.head_size,
)
else:
kv_cache_shape = self.attn_backend.get_kv_cache_shape(
num_blocks, kv_cache_spec.block_size, kv_cache_spec.num_kv_heads, kv_cache_spec.head_size
)
dtype = kv_cache_spec.dtype
k_shape = kv_cache_shape[1:]
v_shape = k_shape
k_cache = torch_npu.empty_with_format(
size=k_shape, dtype=dtype, device=self.device, acl_format=self._acl_format
)
v_cache = torch_npu.empty_with_format(
size=v_shape, dtype=dtype, device=self.device, acl_format=self._acl_format
)
kv_caches[layer_name] = (k_cache, v_cache)
elif isinstance(kv_cache_spec, MambaSpec):
tensor_size = kv_cache_sizes[layer_name]
dtype = kv_cache_spec.dtype
tensor_element_size = torch.tensor([], dtype=dtype).element_size()
raw_tensor = torch.zeros(tensor_size // tensor_element_size, dtype=dtype, device=self.device)
assert tensor_size is not None
assert tensor_size % kv_cache_spec.page_size_bytes == 0
num_blocks = tensor_size // kv_cache_spec.page_size_bytes
assert num_blocks >= kv_cache_config.num_blocks
state_tensors = []
target_idx = 0
start_idx = 0
for shape, dtype in zip(kv_cache_spec.shapes, kv_cache_spec.dtypes):
# normally, there is conv state and ssm state in this loop. And there is only
# a conv state in some special models.
target_shape = (num_blocks, *shape)
target_idx += torch.prod(torch.tensor(target_shape)).item()
tensor = raw_tensor[start_idx:target_idx].view(target_shape)
start_idx = target_idx
state_tensors.append(tensor)
kv_caches[layer_name] = state_tensors
else:
raise ValueError("Unknown KV cache spec type.")
return kv_caches
# Override this function because of tensor.copy_(other) accuracy issue.
# TODO: This override will be removed after tensor.copy_(other) accuracy issue is resolved.
def _prepare_input_ids(
self,
scheduler_output: SchedulerOutput,
total_num_scheduled_tokens: int,
cu_num_tokens: np.ndarray,
) -> None:
"""Prepare the input IDs for the current batch.
Carefully handles the `prev_sampled_token_ids` which can be cached
from the previous engine iteration, in which case those tokens on the
GPU need to be copied into the corresponding slots into input_ids."""
if self.input_batch.prev_sampled_token_ids is None:
# Normal scheduling case
self.input_ids.copy_to_gpu(total_num_scheduled_tokens)
if self.enable_prompt_embeds:
self.inputs_embeds.copy_to_gpu(total_num_scheduled_tokens)
self.is_token_ids.copy_to_gpu(total_num_scheduled_tokens)
return
# Async scheduling case, where some decode requests from the previous
# iteration won't have entries in input_ids_cpu and need to be copied
# on the NPU from prev_sampled_token_ids.
prev_req_id_to_index = self.input_batch.prev_req_id_to_index
assert prev_req_id_to_index is not None
sample_flattened_indices: list[int] = []
spec_flattened_indices: list[int] = []
prev_common_req_indices: list[int] = []
prev_draft_token_indices: list[int] = []
indices_match = True
max_flattened_index = -1
total_num_spec_tokens = 0
scheduled_spec_tokens = scheduler_output.scheduled_spec_decode_tokens
for req_id, cur_index in self.input_batch.req_id_to_index.items():
if (prev_index := prev_req_id_to_index.get(req_id)) is not None:
prev_common_req_indices.append(prev_index)
draft_len = len(scheduled_spec_tokens.get(req_id, ()))
total_num_spec_tokens += draft_len
flattened_index = cu_num_tokens[cur_index].item() - 1
sample_flattened_indices.append(flattened_index - draft_len)
spec_flattened_indices.extend(range(flattened_index - draft_len + 1, flattened_index + 1))
start = prev_index * self.num_spec_tokens
prev_draft_token_indices.extend(range(start, start + draft_len))
indices_match &= prev_index == flattened_index
max_flattened_index = max(max_flattened_index, flattened_index)
num_common_tokens = len(sample_flattened_indices)
total_without_spec = total_num_scheduled_tokens - total_num_spec_tokens
if num_common_tokens < total_without_spec:
self.input_ids.copy_to_gpu(total_num_scheduled_tokens)
if self.enable_prompt_embeds:
self.inputs_embeds.copy_to_gpu(total_num_scheduled_tokens)
self.is_token_ids.copy_to_gpu(total_num_scheduled_tokens)
if num_common_tokens == 0:
return
if indices_match and max_flattened_index == (num_common_tokens - 1):
# NOTE: Override the copy_ function here
indices = torch.arange(num_common_tokens, device=self.input_ids.gpu.device)
source = self.input_batch.prev_sampled_token_ids[:num_common_tokens, 0]
self.input_ids.gpu.index_copy_(0, indices, source)
if self.enable_prompt_embeds:
self.is_token_ids.gpu[:num_common_tokens] = True
return
# Upload the index tensors asynchronously so the scatter can be non-blocking.
sampled_tokens_index_tensor = torch.tensor(
sample_flattened_indices, dtype=torch.int64, pin_memory=self.pin_memory
).to(self.device, non_blocking=True)
prev_common_req_indices_tensor = torch.tensor(
prev_common_req_indices, dtype=torch.int64, pin_memory=self.pin_memory
).to(self.device, non_blocking=True)
self.input_ids.gpu.scatter_(
dim=0,
index=sampled_tokens_index_tensor,
src=self.input_batch.prev_sampled_token_ids[prev_common_req_indices_tensor, 0],
)
# Scatter the draft tokens after the sampled tokens are scattered.
if self._draft_token_ids is None or not spec_flattened_indices:
return
assert isinstance(self._draft_token_ids, torch.Tensor)
draft_tokens_index_tensor = torch.tensor(
spec_flattened_indices, dtype=torch.int64, pin_memory=self.pin_memory
).to(self.device, non_blocking=True)
prev_draft_token_indices_tensor = torch.tensor(
prev_draft_token_indices, dtype=torch.int64, pin_memory=self.pin_memory
).to(self.device, non_blocking=True)
draft_token_ids = self._draft_token_ids.to(dtype=torch.int32)
self.input_ids.gpu.scatter_(
dim=0,
index=draft_tokens_index_tensor,
src=draft_token_ids.flatten()[prev_draft_token_indices_tensor],
)