[Feature] adapt to uva buffer and main2main (#6657)

### What this PR does / why we need it?
vllm model runner v2 use uva buffer to prepare input data, but npu
doesn't support uva yet, this pr implement a uvawrapper class to mimic
gpu's uva backend. what's more, this pr make some modifications to adapt
to the newer main branch.

### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
- vLLM main:
13397841ab

---------

Signed-off-by: Ronald1995 <ronaldautomobile@163.com>
This commit is contained in:
Ronald
2026-02-12 10:36:31 +08:00
committed by GitHub
parent 56269eae0e
commit f1ffb5fb19
14 changed files with 407 additions and 179 deletions

View File

@@ -263,3 +263,12 @@
# https://gitcode.com/Ascend/torchair/pull/2575
# Future Plan:
# Remove this patch when the PTA version used by vllm-ascend has been upgraded.
# ** 14. File: worker/patch_v2_uva.py**
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# 1. `vllm.v1.worker.gpu.states.UvaBuffer`
# Why:
# ASCEND NPUs do not support UVA yet, so we need to wrap it in vLLM.
# How
# make UvaBuffer a dummy class, mimic the interface of vllm UvaBuffer.
# Future Plan:
# Remove this patch when NPU support UVA.

View File

@@ -31,6 +31,7 @@ import vllm_ascend.patch.worker.patch_qwen3_next # noqa
import vllm_ascend.patch.worker.patch_qwen3_next_mtp # noqa
import vllm_ascend.patch.worker.patch_rejection_sampler # noqa
import vllm_ascend.patch.worker.patch_qwen3_next # noqa
import vllm_ascend.patch.worker.patch_v2_egale # noqa
import vllm_ascend.patch.worker.patch_v2_eagle # noqa
import vllm_ascend.patch.worker.patch_v2_uva # noqa
import vllm_ascend.patch.worker.patch_huanyuan_vl # noqa
import vllm_ascend.patch.worker.patch_npugraph_ex_triton # noqa

View File

@@ -16,10 +16,9 @@
# limitations under the License.
# This file is a part of the vllm-ascend project.
#
import numpy as np
import torch
import vllm
from vllm.v1.sample.metadata import SamplingMetadata
from vllm.v1.worker.gpu.attn_utils import build_slot_mappings_by_layer
from vllm.v1.worker.gpu.input_batch import InputBatch
from vllm.v1.worker.gpu.sample.gumbel import gumbel_sample
from vllm.v1.worker.gpu.spec_decode.eagle import prepare_eagle_decode, prepare_eagle_inputs
@@ -31,7 +30,6 @@ from vllm_ascend.worker.v2.attn_utils import build_attn_metadata
def propose(
self,
input_batch: InputBatch,
sampling_metadata: SamplingMetadata,
# [num_tokens, hidden_size]
last_hidden_states: torch.Tensor,
# num_layers x [num_tokens, hidden_size]
@@ -40,10 +38,14 @@ def propose(
num_sampled: torch.Tensor,
# [num_reqs]
num_rejected: torch.Tensor,
# [num_reqs]
# [max_num_reqs]
last_sampled: torch.Tensor,
# [num_reqs]
# [max_num_reqs]
next_prefill_tokens: torch.Tensor,
# [max_num_reqs]
temperature: torch.Tensor,
# [max_num_reqs]
seeds: torch.Tensor,
) -> torch.Tensor:
# NOTE(woosuk): To avoid CPU-GPU synchronization without CPU knowing the
# number of rejected tokens, we maintain the size of eagle's input_ids and
@@ -74,13 +76,13 @@ def propose(
last_hidden_states, hidden_states = self.run_model(
num_tokens,
input_batch.attn_metadata,
input_batch.slot_mappings,
num_tokens_across_dp=None, # FIXME
)
sample_hidden_states = last_hidden_states[last_token_indices]
logits = self.model.compute_logits(sample_hidden_states)
num_reqs = input_batch.num_reqs
cu_num_logits = input_batch.cu_num_logits[:num_reqs]
# NOTE(woosuk): For draft sampling, we only consider the temperature
# and ignore the other sampling parameters such as top_k and top_p,
# for simplicity and performance.
@@ -89,16 +91,23 @@ def propose(
# NOTE(Ronald1995): torch.gather will pollute the cache such as self.input_buffers.positions
# the bug is reported to huawei CANN team, but not fixed yet.
# So we clone the tensors before calling torch.gather to avoid the issue.
temperature = self.temperature[:num_reqs].clone()
seeds = self.seeds[:num_reqs].clone()
idx_mapping = self.idx_mapping[:num_reqs]
idx_mapping.copy_(input_batch.idx_mapping)
self.temperature.copy_(temperature)
self.seeds.copy_(seeds)
pos = self.input_buffers.positions[:num_reqs].clone()
# Gather the values and copy them to the pre-allocated buffers.
torch.gather(sampling_metadata.temperature, 0, cu_num_logits, out=temperature)
torch.gather(sampling_metadata.seeds, 0, cu_num_logits, out=seeds)
torch.gather(input_batch.positions, 0, last_token_indices, out=pos)
# NOTE(woosuk): We must add 1 to the positions to match the Gumbel noise
# used for draft and target sampling.
draft_tokens = gumbel_sample(logits, temperature, seeds, pos + 1, apply_temperature=True)
draft_tokens = gumbel_sample(
logits,
idx_mapping,
self.temperature,
self.seeds,
pos + 1,
apply_temperature=True,
)
if self.num_speculative_steps == 1:
# Early exit.
return draft_tokens.view(-1, 1)
@@ -117,9 +126,12 @@ def propose(
self.max_model_len,
self.max_num_reqs,
)
query_start_loc = self.input_buffers.query_start_loc
query_start_loc_gpu = query_start_loc.gpu[: num_reqs + 1]
slot_mappings = self.block_tables.compute_slot_mappings(query_start_loc_gpu, pos)
query_start_loc = self.input_buffers.query_start_loc[: num_reqs + 1]
slot_mappings = self.block_tables.compute_slot_mappings(
idx_mapping,
query_start_loc,
pos,
)
cudagraph_size = self.cudagraph_manager.get_cudagraph_size(num_reqs)
if cudagraph_size is not None:
@@ -128,10 +140,8 @@ def propose(
return self.draft_tokens[:num_reqs]
# Run eager mode.
query_start_loc.np[: num_reqs + 1] = np.arange(num_reqs + 1)
query_start_loc_cpu = query_start_loc.cpu[: num_reqs + 1]
query_start_loc_cpu = torch.arange(num_reqs + 1, dtype=torch.int32, device="cpu")
# HACK(woosuk)
seq_lens_np = np.full(num_reqs, self.max_model_len, dtype=np.int32)
block_tables = [x[:num_reqs] for x in self.block_tables.input_block_tables]
# FIXME(woosuk): This is UNSAFE!!
@@ -139,16 +149,22 @@ def propose(
attn_metadata_builders=self.attn_metadata_builders,
num_reqs=num_reqs,
num_tokens=num_reqs,
query_start_loc_gpu=query_start_loc_gpu,
query_start_loc_gpu=query_start_loc,
query_start_loc_cpu=query_start_loc_cpu,
max_query_len=1,
seq_lens=self.input_buffers.seq_lens[:num_reqs],
seq_lens_np=seq_lens_np,
num_computed_tokens_cpu=None, # FIXME
max_seq_len=self.max_model_len,
block_tables=block_tables,
slot_mappings=slot_mappings,
kv_cache_config=self.kv_cache_config,
)
self.generate_draft(num_reqs, attn_metadata, num_tokens_across_dp=None) # FIXME
slot_mappings_by_layer = build_slot_mappings_by_layer(slot_mappings, self.kv_cache_config)
self.generate_draft(
num_reqs,
attn_metadata,
slot_mappings_by_layer,
num_tokens_across_dp=None,
) # FIXME
return self.draft_tokens[:num_reqs]

View File

@@ -0,0 +1,125 @@
# Adapt from https://github.com/vllm-project/vllm/blob/main/vllm/v1/worker/gpu/block_table.py
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# 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 collections.abc import Callable, Sequence
import numpy as np
import torch
import vllm.v1.worker.gpu.buffer_utils
def get_row_indices_from_key(key: int | slice | tuple, dim_size: int) -> set[int]:
"""get the set of row indices involved in the given key."""
if isinstance(key, int):
# parse index such as np[1]
key = key if key >= 0 else dim_size + key
# handle negative index
if key < 0 or key >= dim_size:
raise IndexError(f"row index {key} out of [0, {dim_size})")
return {key}
elif isinstance(key, slice):
# parse slice such as np[1:3]
start, stop, step = key.indices(dim_size)
return set(range(start, stop, step))
elif isinstance(key, tuple):
# parse row slice such as np[1,:100]
if len(key) == 0:
return set(range(dim_size))
return get_row_indices_from_key(key[0], dim_size)
else:
# for other types such as list/ndarray, we return all rows.
return set(range(dim_size))
class MonitoredNumPyArray:
"""A wrapper around a NumPy array that monitors modifications."""
def __init__(self, array: np.ndarray, callback: Callable):
self._array = array
self._callback = callback
def __setitem__(self, key, value):
self._array[key] = value
dim_size = self._array.shape[0]
row_indices = get_row_indices_from_key(key, dim_size)
for row in row_indices:
self._callback(row)
def __getitem__(self, key):
return self._array[key]
def __getattr__(self, name):
return getattr(self._array, name)
class MonitoredTorchTensor:
"""A wrapper around a torch tensor that monitors modifications."""
def __init__(self, tensor: torch.Tensor, callback: Callable):
self._tensor = tensor
self._callback = callback
def __setitem__(self, key, value):
self._tensor[key] = value
dim_size = self._tensor.size(0)
row_indices = get_row_indices_from_key(key, dim_size)
for row in row_indices:
self._callback(row)
def __getitem__(self, key):
return self._tensor[key]
def __getattr__(self, name):
return getattr(self._tensor, name)
class UvaBufferWrapper:
"""Ascend NPU doesn't support UVA tensors directly. This is a wrapper class
that provides CPU and NPU views of a UVA tensor."""
def __init__(self, size: int | Sequence[int], dtype: torch.dtype):
self._cpu: torch.Tensor = torch.zeros(size, dtype=dtype, device="cpu", pin_memory=True)
self._np = self._cpu.numpy()
self._uva: torch.Tensor = torch.zeros_like(self._cpu, device="npu")
self._modified_indices: set[int] = set()
def _mark_cpu_modified(self, key: int):
self._modified_indices.add(key)
@property
def cpu(self):
return MonitoredTorchTensor(self._cpu, self._mark_cpu_modified)
@property
def np(self):
return MonitoredNumPyArray(self._np, self._mark_cpu_modified)
@property
def uva(self):
"""Get the device data of the buffer."""
if self._modified_indices:
# Sort for better memory access locality
dirty_rows = sorted(self._modified_indices)
# can't use copy_ method, because copy_ for index tensor
# will malloc new memory.
self._uva[dirty_rows] = self._cpu[dirty_rows].to(device="npu", non_blocking=True)
self._modified_indices.clear()
return self._uva
vllm.v1.worker.gpu.buffer_utils.UvaBuffer = UvaBufferWrapper

View File

@@ -551,7 +551,7 @@ class NPUPlatform(Platform):
vllm_config: VllmConfig,
dp_metadata,
virtual_engine: int = 0,
num_tokens: int | None = None,
num_tokens: int = 0,
num_tokens_across_dp: torch.Tensor | None = None,
cudagraph_runtime_mode=None,
batch_descriptor=None,
@@ -601,10 +601,6 @@ class NPUPlatform(Platform):
if not envs_vllm.VLLM_USE_V2_MODEL_RUNNER:
return {}
num_actual_tokens = list(attn_metadata.values())[0].num_actual_tokens
if num_tokens is None:
num_tokens = num_actual_tokens
moe_comm_type = select_moe_comm_method(
num_tokens,
vllm_config,
@@ -636,10 +632,13 @@ class NPUPlatform(Platform):
# TODO(Levi-JQ): another PR to normalize the enabling logic for sp/fc2
flashcomm_v2_enabled = flashcomm2_enable() and tp_world_size > 1 and num_tokens is not None
pad_size = 0
pad_size = None
padded_length = None
if sp_enabled or flashcomm_v2_enabled:
pad_size = (tp_world_size - (num_tokens % tp_world_size)) % tp_world_size
if num_tokens is None and attn_metadata is not None:
num_tokens = list(attn_metadata.values())[0].num_actual_tokens
dp_world_size = get_dp_group().world_size
if dp_world_size > 1 and dp_metadata is not None:
max_tokens_across_dp = dp_metadata.max_tokens_across_dp_cpu.item()
@@ -648,8 +647,9 @@ class NPUPlatform(Platform):
pad_size = padded_length - num_tokens
else:
max_tokens_across_dp = num_tokens
mc2_mask = None
if num_tokens is not None:
num_actual_tokens = num_tokens
# NOTE: token num which need to pad to when mc2
padded_num_tokens = math.ceil(max_tokens_across_dp / tp_world_size) * tp_world_size
reserved_mc2_mask = get_mc2_mask()

View File

@@ -4,3 +4,5 @@ This directory contains the new model runner which is under active development.
please see [Model Runner V2](https://github.com/vllm-project/vllm-ascend/issues/5208)
to get specific plans.
supported vllm version: main@1339784

View File

@@ -35,9 +35,14 @@ from vllm_ascend.worker.v2.utils import torch_cuda_wrapper
class AclGraphManager(CudaGraphManager):
"""ACL Graph Manager for Ascend NPUs."""
def __init__(self, vllm_config: VllmConfig, device: torch.device):
def __init__(
self,
vllm_config: VllmConfig,
use_mrope: bool,
device: torch.device,
):
with torch_cuda_wrapper():
super().__init__(vllm_config, device)
super().__init__(vllm_config, use_mrope, device)
def capture_graph(
self,

View File

@@ -42,17 +42,21 @@ def get_attn_mask_builder(device: torch.device):
def build_attn_metadata(
*,
attn_metadata_builders: list[AttentionMetadataBuilder],
num_reqs: int,
num_tokens: int,
query_start_loc_gpu: torch.Tensor,
query_start_loc_cpu: torch.Tensor,
max_query_len: int,
seq_lens: torch.Tensor,
seq_lens_np: np.ndarray,
num_computed_tokens_cpu: torch.Tensor | None,
max_seq_len: int,
block_tables: Sequence[torch.Tensor],
slot_mappings: torch.Tensor,
kv_cache_config: KVCacheConfig,
# extra attributes for ascend npus.
seq_lens_np: np.ndarray | None = None,
num_computed_tokens_cpu: torch.Tensor | None = None,
positions: torch.Tensor | None = None,
attn_state: Any | None = None,
graph_pad_size: int = -1,
@@ -61,9 +65,13 @@ def build_attn_metadata(
) -> dict[str, Any]:
"""Build attention metadata for Ascend NPUs."""
# TODO(Ronald1995): optimize AscendCommonAttentionMetadata.
max_query_len = int(query_start_loc_cpu.max())
seq_lens_cpu = torch.from_numpy(seq_lens_np)
max_seq_len = int(seq_lens_cpu.max())
# seq_lens_np is used for ascend npus, it maybe None in spec_decode case,
# we fill it with max_seq_len in case `attn_metadata_builder.build` raise
# an error.
if seq_lens_np is None:
seq_lens_np = np.full(num_reqs, max_seq_len, dtype=np.int32)
seq_lens_cpu = torch.from_numpy(seq_lens_np)[:num_reqs]
# torch_npu._reshape_and_cache operator requires slot_mappings to
# be torch.int32.
slot_mappings = slot_mappings.to(torch.int32)
@@ -77,7 +85,7 @@ def build_attn_metadata(
common_attn_metadata = AscendCommonAttentionMetadata(
query_start_loc=query_start_loc_gpu,
query_start_loc_cpu=query_start_loc_cpu,
seq_lens_cpu=seq_lens_cpu[:num_reqs],
seq_lens_cpu=seq_lens_cpu,
seq_lens=seq_lens[:num_reqs],
num_reqs=num_reqs,
num_actual_tokens=num_tokens,

View File

@@ -16,10 +16,11 @@
# limitations under the License.
# This file is a part of the vllm-ascend project.
#
from dataclasses import asdict, dataclass
import numpy as np
import torch
from vllm.v1.worker.gpu.input_batch import InputBuffers
from vllm.v1.worker.gpu.input_batch import InputBatch, InputBuffers
class AscendInputBuffers(InputBuffers):
@@ -29,20 +30,12 @@ class AscendInputBuffers(InputBuffers):
self,
max_num_reqs: int,
max_num_tokens: int,
inputs_embeds_size: int,
vocab_size: int,
dtype: torch.dtype,
device: torch.device,
pin_memory: bool,
):
super().__init__(
max_num_reqs,
max_num_tokens,
inputs_embeds_size,
vocab_size,
dtype,
device,
pin_memory,
)
# Create seq_lens_cpu and seq_lens_np.
# npu's attention backend still needs seq_lens on CPU side.
@@ -54,3 +47,36 @@ class AscendInputBuffers(InputBuffers):
# seq_len_np and seq_lens_cpu share the same memory.
# define seq_lens_np for easier calculation with numpy.
self.seq_lens_np: np.ndarray = self.seq_lens_cpu.numpy()
@dataclass
class AscendInputBatch(InputBatch):
"""Input batch for Ascend NPUs."""
# Create seq_lens_np.
# npu's attention backend still needs seq_lens on CPU side.
seq_lens_np: np.ndarray
@classmethod
def make_dummy(
cls,
num_reqs: int,
num_tokens: int,
input_buffers: AscendInputBuffers,
device: torch.device,
) -> "AscendInputBatch":
"""Override the make_dummy method to calculate seq_lens_np."""
input_batch = InputBatch.make_dummy(
num_reqs,
num_tokens,
input_buffers,
device,
)
# seq_len equals to query_len
input_buffers.seq_lens_np[:num_reqs] = num_tokens // num_reqs
input_buffers.seq_lens_np[num_reqs - 1] += num_tokens % num_reqs
# Pad for full CUDA graph mode.
input_buffers.seq_lens_np[num_reqs:] = 0
seq_lens_np = input_buffers.seq_lens_np[:num_reqs]
input_batch.seq_lens_np = seq_lens_np
return cls(**asdict(input_batch), seq_lens_np=seq_lens_np)

View File

@@ -22,9 +22,11 @@ import torch
from vllm.config import VllmConfig
from vllm.logger import init_logger
from vllm.v1.core.sched.output import SchedulerOutput
from vllm.v1.worker.gpu.attn_utils import build_slot_mappings_by_layer
from vllm.v1.worker.gpu.buffer_utils import async_copy_to_gpu
from vllm.v1.worker.gpu.input_batch import (
InputBatch,
combine_sampled_and_draft_tokens,
expand_idx_mapping,
prepare_pos_seq_lens,
prepare_prefill_inputs,
)
@@ -32,11 +34,11 @@ from vllm.v1.worker.gpu.model_runner import GPUModelRunner
from vllm_ascend.worker.v2.aclgraph_utils import AclGraphManager
from vllm_ascend.worker.v2.attn_utils import build_attn_metadata, build_attn_state
from vllm_ascend.worker.v2.input_batch import AscendInputBuffers
from vllm_ascend.worker.v2.input_batch import AscendInputBatch, AscendInputBuffers
from vllm_ascend.worker.v2.sample.sampler import AscendSampler
from vllm_ascend.worker.v2.spec_decode import init_speculator
from vllm_ascend.worker.v2.spec_decode.eagle import AscendEagleSpeculator
from vllm_ascend.worker.v2.states import AscendRequestState, uva_wrapper
from vllm_ascend.worker.v2.states import AscendRequestState
from vllm_ascend.worker.v2.utils import torch_cuda_wrapper
logger = init_logger(__name__)
@@ -46,7 +48,7 @@ class NPUModelRunner(GPUModelRunner):
"""Model runner for Ascend NPUs."""
def __init__(self, vllm_config: VllmConfig, device: torch.device):
with torch_cuda_wrapper(), uva_wrapper():
with torch_cuda_wrapper():
super().__init__(vllm_config, device)
# because we will override these attribute, delete these attribute to
@@ -59,8 +61,9 @@ class NPUModelRunner(GPUModelRunner):
# NPU specific initializations can be added below.
self.cudagraph_manager: AclGraphManager = AclGraphManager(
vllm_config,
device,
self.vllm_config,
self.uses_mrope,
self.device,
)
# we define AscendEagleSpeculator in vllm_ascend.worker.v2.spec_decode.eagle
@@ -79,23 +82,22 @@ class NPUModelRunner(GPUModelRunner):
num_speculative_steps=self.num_speculative_steps,
vocab_size=self.vocab_size,
device=self.device,
pin_memory=self.pin_memory,
)
# AscendInputBuffers has extra `seq_lens_cpu` attribute.
# so reinitialize input_buffers here.
self.input_buffers: AscendInputBuffers = AscendInputBuffers(
max_num_reqs=self.max_num_reqs,
max_num_tokens=self.max_num_tokens,
inputs_embeds_size=self.inputs_embeds_size,
vocab_size=self.vocab_size,
dtype=self.dtype,
device=self.device,
pin_memory=self.pin_memory,
)
# we need to adjust triton operators in sampler,
# so reinitialize sampler here.
self.sampler: AscendSampler = AscendSampler(
max_num_reqs=self.max_num_reqs,
vocab_size=self.vocab_size,
device=self.device,
logprobs_mode=self.model_config.logprobs_mode,
num_speculative_tokens=self.num_speculative_steps + 1,
)
# we need to copy num_computed_tokens back to cpu to help
@@ -108,32 +110,31 @@ class NPUModelRunner(GPUModelRunner):
self.max_num_reqs,
dtype=torch.int32,
device="cpu",
pin_memory=self.pin_memory,
pin_memory=True,
)
def prepare_inputs(
self,
scheduler_output: SchedulerOutput,
num_tokens_after_padding: int,
) -> InputBatch:
) -> AscendInputBatch:
"""Override GPUModelRunner.prepare_inputs for Ascend NPUs.
npu attention backends need seq_lens_cpu to work.
so we need to prepare seq_lens_cpu here.
"""
num_tokens = scheduler_output.total_num_scheduled_tokens
assert num_tokens > 0
num_reqs = len(scheduler_output.num_scheduled_tokens)
num_tokens_per_req = scheduler_output.num_scheduled_tokens
num_reqs = len(num_tokens_per_req)
# Decode first, then prefill.
# batch_idx -> req_id
req_ids = sorted(
scheduler_output.num_scheduled_tokens.keys(),
key=lambda k: scheduler_output.num_scheduled_tokens[k],
)
req_ids = sorted(num_tokens_per_req, key=num_tokens_per_req.get) # type: ignore
self._update_seq_lens_cpu(scheduler_output, req_ids)
num_scheduled_tokens = np.array([scheduler_output.num_scheduled_tokens[i] for i in req_ids], dtype=np.int32)
numtoks_iter = map(num_tokens_per_req.get, req_ids)
num_scheduled_tokens = np.fromiter(numtoks_iter, dtype=np.int32, count=num_reqs)
num_valid_tokens = num_scheduled_tokens
if scheduler_output.scheduled_spec_decode_tokens:
num_valid_tokens = np.array(
@@ -150,81 +151,94 @@ class NPUModelRunner(GPUModelRunner):
num_scheduled_tokens,
num_valid_tokens,
)
idx_mapping_list = [self.req_states.req_id_to_index[req_id] for req_id in req_ids]
idx_mapping = self.input_buffers.idx_mapping
idx_mapping.np[:num_reqs] = idx_mapping_list
idx_mapping_np = idx_mapping.np[:num_reqs]
idx_mapping_cpu = idx_mapping.cpu[:num_reqs]
idx_mapping_npu = idx_mapping.copy_to_gpu(num_reqs)
idx_mapping_iter = map(self.req_states.req_id_to_index.get, req_ids)
idx_mapping_np = np.fromiter(idx_mapping_iter, dtype=np.int32, count=num_reqs)
idx_mapping_cpu = torch.from_numpy(idx_mapping_np)
idx_mapping = async_copy_to_gpu(idx_mapping_cpu, device=self.device)
# Get the number of draft tokens for each request.
if not scheduler_output.scheduled_spec_decode_tokens:
draft_tokens = scheduler_output.scheduled_spec_decode_tokens
if not draft_tokens:
# No draft token scheduled (common case).
total_num_draft_tokens = 0
total_num_logits = num_reqs
cu_num_logits_np = np.arange(num_reqs + 1, dtype=np.int32)
cu_num_logits = torch.arange(num_reqs + 1, device=self.device, dtype=torch.int32)
expanded_idx_mapping = idx_mapping
expanded_local_pos = torch.zeros(num_reqs, dtype=torch.int32, device=self.device)
else:
draft_tokens = scheduler_output.scheduled_spec_decode_tokens
num_draft_tokens = np.array(
[len(draft_tokens[req_id]) if req_id in draft_tokens else 0 for req_id in req_ids],
[len(draft_tokens.get(req_id, ())) for req_id in req_ids],
dtype=np.int32,
)
total_num_draft_tokens = int(num_draft_tokens.sum())
total_num_logits = num_reqs + total_num_draft_tokens
np.cumsum(
num_draft_tokens + 1,
out=self.input_buffers.cu_num_logits.np[1 : num_reqs + 1],
num_logits = num_draft_tokens + 1
cu_num_logits_np = np.empty(num_reqs + 1, dtype=np.int32)
cu_num_logits_np[0] = 0
np.cumsum(num_logits, out=cu_num_logits_np[1:])
cu_num_logits = async_copy_to_gpu(cu_num_logits_np, device=self.device)
max_expand_len = self.num_speculative_steps + 1
expanded_idx_mapping, expanded_local_pos = expand_idx_mapping(
idx_mapping, total_num_logits, cu_num_logits, max_expand_len
)
cu_num_logits = self.input_buffers.cu_num_logits.copy_to_gpu(num_reqs + 1)
# Block tables: num_kv_cache_groups x [num_reqs, max_num_blocks]
block_tables = self.block_tables.gather_block_tables(idx_mapping_npu)
block_tables = self.block_tables.gather_block_tables(idx_mapping)
# Get query_start_loc.
np.cumsum(
num_scheduled_tokens,
out=self.input_buffers.query_start_loc.np[1 : num_reqs + 1],
)
query_start_loc_np = np.empty(self.max_num_reqs + 1, dtype=np.int32)
query_start_loc_np[0] = 0
np.cumsum(num_scheduled_tokens, out=query_start_loc_np[1 : num_reqs + 1])
# Pad for full CUDA graph mode.
# Some attention backends like FA3 require query_start_loc to be non-decreasing.
self.input_buffers.query_start_loc.np[num_reqs + 1 :] = num_tokens
self.input_buffers.query_start_loc.copy_to_gpu()
query_start_loc_gpu = self.input_buffers.query_start_loc.gpu[: num_reqs + 1]
query_start_loc_cpu = self.input_buffers.query_start_loc.cpu[: num_reqs + 1]
query_start_loc_np = self.input_buffers.query_start_loc.np[: num_reqs + 1]
query_start_loc_np[num_reqs + 1 :] = num_tokens
async_copy_to_gpu(query_start_loc_np, out=self.input_buffers.query_start_loc)
query_start_loc_np = query_start_loc_np[: num_reqs + 1]
query_start_loc_cpu = torch.from_numpy(query_start_loc_np)
query_start_loc = self.input_buffers.query_start_loc[: num_reqs + 1]
max_query_len = num_scheduled_tokens.max().item()
# Get prefill tokens.
prepare_prefill_inputs(
self.input_buffers.input_ids,
self.req_states.next_prefill_tokens,
idx_mapping_npu,
query_start_loc_gpu,
# use prefill_token_ids.copy_to_gpu() because npu doesn't
# support uva buffer.
self.req_states.prefill_token_ids.copy_to_gpu(),
idx_mapping,
query_start_loc,
self.req_states.prefill_token_ids.gpu,
self.req_states.prefill_len.gpu,
self.req_states.num_computed_tokens,
self.req_states.num_computed_tokens.gpu,
)
# Prepare positions and seq_lens.
prepare_pos_seq_lens(
idx_mapping_npu,
query_start_loc_gpu,
self.req_states.num_computed_tokens,
idx_mapping,
query_start_loc,
self.req_states.num_computed_tokens.gpu,
self.input_buffers.positions,
self.input_buffers.seq_lens,
)
seq_lens = self.input_buffers.seq_lens[:num_reqs]
# Prepare M-RoPE positions.
if self.uses_mrope:
self.mrope_states.prepare_mrope_positions(
idx_mapping,
query_start_loc,
self.req_states.prefill_len.gpu,
self.req_states.num_computed_tokens.gpu,
)
# Some input token ids are directly read from the last sampled tokens
# and draft tokens. Also, get the logits indices to sample tokens from.
logits_indices = combine_sampled_and_draft_tokens(
self.input_buffers.input_ids,
idx_mapping_npu,
idx_mapping,
self.req_states.last_sampled_tokens,
query_start_loc_gpu,
query_start_loc,
seq_lens,
self.req_states.prefill_len.gpu,
self.req_states.draft_tokens,
@@ -234,9 +248,10 @@ class NPUModelRunner(GPUModelRunner):
# Compute slot mappings: [num_kv_cache_groups, num_tokens]
slot_mappings = self.block_tables.compute_slot_mappings(
query_start_loc_gpu, self.input_buffers.positions[:num_tokens]
idx_mapping, query_start_loc, self.input_buffers.positions[:num_tokens]
)
# Layer name -> slot mapping.
slot_mappings_by_layer = build_slot_mappings_by_layer(slot_mappings, self.kv_cache_config)
# Layer name -> attention metadata.
# TODO(Ronald1995): try to add a new method `build_attn_metadata` in
# vllm gpu_model_runner_v2, maybe we don't overwrite `prepare_inputs`
@@ -245,37 +260,51 @@ class NPUModelRunner(GPUModelRunner):
attn_metadata_builders=self.attn_metadata_builders,
num_reqs=num_reqs,
num_tokens=num_tokens,
query_start_loc_gpu=query_start_loc_gpu,
query_start_loc_gpu=query_start_loc,
query_start_loc_cpu=query_start_loc_cpu,
max_query_len=max_query_len,
seq_lens=self.input_buffers.seq_lens,
seq_lens_np=self.input_buffers.seq_lens_np,
num_computed_tokens_cpu=self.req_states.num_computed_tokens_cpu[idx_mapping_cpu],
max_seq_len=self.max_model_len,
block_tables=block_tables,
slot_mappings=slot_mappings,
kv_cache_config=self.kv_cache_config,
# extra attributes for ascend npus.
seq_lens_np=self.input_buffers.seq_lens_np,
num_computed_tokens_cpu=self.req_states.num_computed_tokens_cpu[idx_mapping_cpu],
attn_state=attn_state,
)
input_ids = self.input_buffers.input_ids[:num_tokens_after_padding]
positions = self.input_buffers.positions[:num_tokens_after_padding]
return InputBatch(
mrope_positions = None
if self.uses_mrope:
mrope_positions = self.mrope_states.mrope_positions
mrope_positions = mrope_positions[:, :num_tokens_after_padding]
return AscendInputBatch(
req_ids=req_ids,
num_reqs=num_reqs,
idx_mapping=idx_mapping_npu,
idx_mapping=idx_mapping,
idx_mapping_np=idx_mapping_np,
expanded_idx_mapping=expanded_idx_mapping,
expanded_local_pos=expanded_local_pos,
num_scheduled_tokens=num_scheduled_tokens,
num_tokens=num_tokens,
num_tokens_after_padding=num_tokens_after_padding,
num_draft_tokens=total_num_draft_tokens,
query_start_loc=query_start_loc_gpu,
query_start_loc=query_start_loc,
query_start_loc_np=query_start_loc_np,
seq_lens=seq_lens,
seq_lens_np=self.input_buffers.seq_lens_np,
input_ids=input_ids,
positions=positions,
mrope_positions=mrope_positions,
inputs_embeds=None,
attn_metadata=attn_metadata,
slot_mappings=slot_mappings_by_layer,
logits_indices=logits_indices,
cu_num_logits=cu_num_logits,
cu_num_logits_np=cu_num_logits_np,
has_structured_output_reqs=scheduler_output.has_structured_output_requests,
seq_lens_np=self.input_buffers.seq_lens_np,
)
def postprocess(
@@ -303,7 +332,7 @@ class NPUModelRunner(GPUModelRunner):
with torch.npu.stream(self.num_computed_tokens_stream):
self.num_computed_tokens_stream.wait_stream(default_stream)
self.num_computed_tokens_cpu.copy_(
self.req_states.num_computed_tokens,
self.req_states.num_computed_tokens.gpu,
non_blocking=True,
)
self.num_computed_tokens_event.record()

View File

@@ -30,6 +30,7 @@ def _gumbel_sample_kernel(
local_max_stride,
logits_ptr,
logits_stride,
idx_mapping_ptr,
seeds_ptr,
pos_ptr,
temp_ptr,
@@ -37,24 +38,26 @@ def _gumbel_sample_kernel(
BLOCK_SIZE: tl.constexpr,
APPLY_TEMPERATURE: tl.constexpr,
):
req_idx = tl.program_id(0)
batch_idx = tl.program_id(0)
req_state_idx = tl.load(idx_mapping_ptr + batch_idx)
block_idx = tl.program_id(1)
block = block_idx * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
mask = block < vocab_size
logits = tl.load(
logits_ptr + req_idx * logits_stride + block,
logits_ptr + batch_idx * logits_stride + block,
mask=mask,
other=float("-inf"),
)
logits = logits.to(tl.float32)
temp = tl.load(temp_ptr + req_idx).to(tl.float32)
temp = tl.load(temp_ptr + req_state_idx).to(tl.float32)
if temp != 0.0:
# Calculate the seed for gumbel noise.
seed = tl.load(seeds_ptr + req_idx)
seed = tl.load(seeds_ptr + req_state_idx)
# NOTE(Ronald1995): change pos's dtype to tl.int32, because triton-ascend's
# compiler doesn't support unint64 of pos arg.
pos = tl.load(pos_ptr + req_idx).to(tl.int32)
pos = tl.load(pos_ptr + batch_idx).to(tl.int32)
gumbel_seed = tl.randint(seed, pos)
# Generate gumbel noise.
@@ -66,7 +69,7 @@ def _gumbel_sample_kernel(
# Apply temperature.
if APPLY_TEMPERATURE:
# NOTE(woosuk): Match the behavior of _penalties_and_temperature_kernel.
# NOTE(woosuk): Match the behavior of _temperature_kernel.
# E.g., if the kernel uses tl.div_rn, we should use tl.div_rn here too.
logits = logits / temp
@@ -76,21 +79,18 @@ def _gumbel_sample_kernel(
idx = tl.argmax(logits, axis=0)
token_id = block_idx * BLOCK_SIZE + idx
value = tl.max(logits, axis=0)
tl.store(local_argmax_ptr + req_idx * local_argmax_stride + block_idx, token_id)
tl.store(local_max_ptr + req_idx * local_max_stride + block_idx, value)
tl.store(local_argmax_ptr + batch_idx * local_argmax_stride + block_idx, token_id)
tl.store(local_max_ptr + batch_idx * local_max_stride + block_idx, value)
def gumbel_sample(
logits: torch.Tensor, # [num_reqs, vocab_size]
idx_mapping: torch.Tensor, # [num_reqs]
temperature: torch.Tensor, # [num_reqs]
seed: torch.Tensor, # [num_reqs]
pos: torch.Tensor, # [num_reqs]
apply_temperature: bool,
) -> torch.Tensor:
"""Override the function because there are some bugs
when _gumbel_sample_kernel runs on npu, we need to make some fixes.
you could read NOTE(Ronald1995) comments to understand.
"""
num_reqs, vocab_size = logits.shape
BLOCK_SIZE = 1024
num_blocks = triton.cdiv(vocab_size, BLOCK_SIZE)
@@ -114,6 +114,7 @@ def gumbel_sample(
local_max.stride(0),
logits,
logits.stride(0),
idx_mapping,
seed,
pos,
temperature,

View File

@@ -14,22 +14,25 @@
# limitations under the License.
# This file is a part of the vllm-ascend project.
#
import numpy as np
import torch
from vllm.v1.sample.metadata import SamplingMetadata
from vllm.v1.sample.ops.topk_topp_sampler import apply_top_k_top_p
from vllm.v1.worker.gpu.sample.gumbel import apply_temperature
from vllm.v1.worker.gpu.sample.min_p import apply_min_p
from vllm.v1.worker.gpu.sample.sampler import Sampler
from vllm_ascend.worker.v2.sample.gumbel import gumbel_sample
from vllm_ascend.worker.v2.sample.penalties import apply_penalties_and_temperature
class AscendSampler(Sampler):
def sample(
self,
logits: torch.Tensor,
sampling_metadata: SamplingMetadata,
idx_mapping: torch.Tensor,
idx_mapping_np: np.ndarray,
pos: torch.Tensor,
input_ids: torch.Tensor,
expanded_local_pos: torch.Tensor,
) -> tuple[torch.Tensor, torch.Tensor]:
"""Override sample method because we need to override triton operators
called in the method.
@@ -37,19 +40,42 @@ class AscendSampler(Sampler):
# Copy logits to a new FP32 tensor.
logits = torch.empty_like(logits, dtype=torch.float32).copy_(logits)
# Apply penalties and temperature in place.
apply_penalties_and_temperature(logits, sampling_metadata)
# Apply min_p in place.
if sampling_metadata.min_p is not None:
apply_min_p(logits, sampling_metadata.min_p)
# Apply top_k and/or top_p. This might return a new tensor.
logits = apply_top_k_top_p(logits, sampling_metadata.top_k, sampling_metadata.top_p)
# Apply logit bias (e.g., allowed_token_ids, min_tokens) in place.
self.logit_bias_state.apply_logit_bias(logits, idx_mapping, idx_mapping_np, pos)
# Apply penalties in place.
self.penalties_state.apply_penalties(
logits,
idx_mapping,
idx_mapping_np,
input_ids,
expanded_local_pos,
self.num_speculative_tokens,
)
# Apply temperature in place.
apply_temperature(logits, idx_mapping, self.sampling_states.temperature.gpu)
# Apply min_p in place if any request has a non-zero min_p.
do_min_p = self.sampling_states.do_min_p(idx_mapping_np)
if do_min_p:
apply_min_p(logits, idx_mapping, self.sampling_states.min_p.gpu)
# Apply top_k and/or top_p. This might return a new tensor.
do_top_k = self.sampling_states.do_top_k(idx_mapping_np)
top_k = self.sampling_states.top_k.gpu[idx_mapping] if do_top_k else None
do_top_p = self.sampling_states.do_top_p(idx_mapping_np)
top_p = self.sampling_states.top_p.gpu[idx_mapping] if do_top_p else None
if do_top_k or do_top_p:
logits = apply_top_k_top_p(logits, top_k, top_p)
# Sample the next token.
sampled = gumbel_sample(
logits,
sampling_metadata.temperature,
sampling_metadata.seeds,
sampling_metadata.pos,
idx_mapping,
self.sampling_states.temperature.gpu,
self.sampling_states.seeds.gpu,
pos,
apply_temperature=False,
)
return sampled, logits

View File

@@ -42,14 +42,23 @@ class AscendEagleSpeculator(EagleSpeculator):
def propose(
self,
input_batch,
sampling_metadata,
last_hidden_states,
aux_hidden_states,
num_sampled,
num_rejected,
last_sampled,
next_prefill_tokens,
input_batch: InputBatch,
# [num_tokens, hidden_size]
last_hidden_states: torch.Tensor,
# num_layers x [num_tokens, hidden_size]
aux_hidden_states: list[torch.Tensor] | None,
# [num_reqs]
num_sampled: torch.Tensor,
# [num_reqs]
num_rejected: torch.Tensor,
# [max_num_reqs]
last_sampled: torch.Tensor,
# [max_num_reqs]
next_prefill_tokens: torch.Tensor,
# [max_num_reqs]
temperature: torch.Tensor,
# [max_num_reqs]
seeds: torch.Tensor,
):
"""Override GPU EagleSpeculator.propose for Ascend NPUs,
because npu attention metadata needs more information,
@@ -62,19 +71,21 @@ class AscendEagleSpeculator(EagleSpeculator):
with build_attn_metadata_wrapper():
return super().propose(
input_batch,
sampling_metadata,
last_hidden_states,
aux_hidden_states,
num_sampled,
num_rejected,
last_sampled,
next_prefill_tokens,
temperature,
seeds,
)
def generate_draft(
self,
num_reqs: int,
attn_metadata: dict[str, Any],
slot_mappings: dict[str, torch.Tensor],
num_tokens_across_dp,
):
"""Override GPU EagleSpeculator.generate_draft for Ascend NPUs, because
@@ -86,6 +97,7 @@ class AscendEagleSpeculator(EagleSpeculator):
return super().generate_draft(
num_reqs,
attn_metadata,
slot_mappings,
num_tokens_across_dp,
)
@@ -94,6 +106,7 @@ class AscendEagleSpeculator(EagleSpeculator):
self,
num_tokens: int,
attn_metadata: dict[str, Any],
slot_mappings: dict[str, torch.Tensor] | None,
num_tokens_across_dp: torch.Tensor | None,
) -> tuple[torch.Tensor, torch.Tensor]:
"""Override GPU EagleSpeculator.run_model for Ascend NPUs, because
@@ -103,6 +116,7 @@ class AscendEagleSpeculator(EagleSpeculator):
last_hidden_states, hidden_states = super().run_model(
num_tokens,
attn_metadata,
slot_mappings,
num_tokens_across_dp,
)

View File

@@ -17,11 +17,7 @@
# This file is a part of the vllm-ascend project.
#
from contextlib import contextmanager
import torch
import vllm
from vllm.v1.utils import CpuGpuBuffer
from vllm.v1.worker.gpu.states import RequestState
@@ -36,7 +32,6 @@ class AscendRequestState(RequestState):
num_speculative_steps: int,
vocab_size: int,
device: torch.device,
pin_memory: bool,
):
super().__init__(
max_num_reqs,
@@ -45,11 +40,7 @@ class AscendRequestState(RequestState):
num_speculative_steps,
vocab_size,
device,
pin_memory,
)
# because we will override these attribute, delete these attribute to
# make sure it's collected by python gc immediately.
del self.prefill_token_ids
# vllm gpu_model_runner_v2 deprecate the seqs_lens_cpu attribute,
# because they think most attention backends do not need it.
# However, Ascend attention backend muse uses seqs_lens_cpu,
@@ -60,11 +51,6 @@ class AscendRequestState(RequestState):
dtype=torch.int32,
device="cpu",
)
# NOTE(Ronald1995): Ascend NPUs do not support UVA yet,
# so we use CpuGpuBuffer to allocate prefill_token_ids buffer.
self.prefill_token_ids: CpuGpuBuffer = self._make_buffer( # type: ignore
(self.max_num_reqs, self.max_model_len), dtype=torch.int32
)
def add_request(
self,
@@ -72,32 +58,12 @@ class AscendRequestState(RequestState):
prompt_len,
prefill_token_ids,
num_computed_tokens,
sampling_params,
lora_request,
):
super().add_request(
req_id,
prompt_len,
prefill_token_ids,
num_computed_tokens,
sampling_params,
lora_request,
)
req_idx = self.req_id_to_index[req_id]
self.num_computed_tokens_cpu[req_idx] = num_computed_tokens
@contextmanager
def uva_wrapper():
"""Context manager to disable UVA for Ascend NPUs."""
class UvaBufferWrapper:
def __init__(self, *args, **kwargs):
pass
try:
# TODO(Ronald1995): rectify this when NPU support uva.
vllm.v1.worker.gpu.states.UvaBuffer = UvaBufferWrapper
yield
finally:
pass