init v0.11.0rc0
This commit is contained in:
312
vllm_ascend/worker/block_table.py
Normal file
312
vllm_ascend/worker/block_table.py
Normal file
@@ -0,0 +1,312 @@
|
||||
from typing import Optional, Union
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
from vllm.distributed import get_dcp_group
|
||||
from vllm.utils import cdiv
|
||||
|
||||
|
||||
class BlockTable:
|
||||
|
||||
def __init__(self,
|
||||
block_size: int,
|
||||
max_num_reqs: int,
|
||||
max_num_blocks_per_req: int,
|
||||
max_num_batched_tokens: int,
|
||||
pin_memory: bool,
|
||||
device: torch.device,
|
||||
kernel_sizes: Union[list[int], None] = None):
|
||||
self.max_num_reqs = max_num_reqs
|
||||
self.max_num_blocks_per_req = max_num_blocks_per_req
|
||||
self.max_num_batched_tokens = max_num_batched_tokens
|
||||
self.pin_memory = pin_memory
|
||||
self.device = device
|
||||
self.physical_block_size = block_size
|
||||
# If kernel_sizes is None or [0], use physical block size (no splitting)
|
||||
if kernel_sizes is None or kernel_sizes == [0]:
|
||||
self.block_size = block_size
|
||||
self.logical_block_size = block_size
|
||||
self.blocks_per_phys_block = 1
|
||||
self.use_hybrid_blocks = False
|
||||
else:
|
||||
# Find the first kernel size that divides physical_block_size evenly
|
||||
selected_kernel_size = None
|
||||
for kernel_size in kernel_sizes:
|
||||
if kernel_size > 0 \
|
||||
and self.physical_block_size % kernel_size == 0:
|
||||
selected_kernel_size = kernel_size
|
||||
break
|
||||
|
||||
if selected_kernel_size is None:
|
||||
raise ValueError(
|
||||
f"None of the kernel sizes {kernel_sizes} can divide "
|
||||
f"physical block size {self.physical_block_size} evenly")
|
||||
|
||||
self.block_size = selected_kernel_size
|
||||
self.logical_block_size = selected_kernel_size
|
||||
self.blocks_per_phys_block = (self.physical_block_size //
|
||||
self.logical_block_size)
|
||||
if self.blocks_per_phys_block > 1:
|
||||
self.use_hybrid_blocks = True
|
||||
else:
|
||||
self.use_hybrid_blocks = False
|
||||
|
||||
if self.use_hybrid_blocks:
|
||||
logical_table_size = (max_num_blocks_per_req *
|
||||
self.blocks_per_phys_block)
|
||||
else:
|
||||
logical_table_size = max_num_blocks_per_req
|
||||
|
||||
self.block_table = torch.zeros(
|
||||
(max_num_reqs, logical_table_size),
|
||||
device=self.device,
|
||||
dtype=torch.int32,
|
||||
)
|
||||
self.block_table_cpu = torch.zeros(
|
||||
(max_num_reqs, logical_table_size),
|
||||
device="cpu",
|
||||
dtype=torch.int32,
|
||||
pin_memory=pin_memory,
|
||||
)
|
||||
self.block_table_np = self.block_table_cpu.numpy()
|
||||
self.num_blocks_per_row = np.zeros(max_num_reqs, dtype=np.int32)
|
||||
|
||||
self.slot_mapping_cpu = torch.zeros(self.max_num_batched_tokens,
|
||||
dtype=torch.int64,
|
||||
device="cpu",
|
||||
pin_memory=self.pin_memory)
|
||||
self.slot_mapping_np = self.slot_mapping_cpu.numpy()
|
||||
self.slot_mapping = torch.zeros(self.max_num_batched_tokens,
|
||||
dtype=torch.int64,
|
||||
device=self.device)
|
||||
try:
|
||||
self.dcp_world_size = get_dcp_group().world_size
|
||||
self.dcp_rank = get_dcp_group().rank_in_group
|
||||
except AssertionError:
|
||||
# DCP might not be initialized in testing
|
||||
self.dcp_world_size = 1
|
||||
self.dcp_rank = 0
|
||||
self.kernel_sizes = kernel_sizes
|
||||
|
||||
def append_row(
|
||||
self,
|
||||
block_ids,
|
||||
row_idx: int,
|
||||
) -> None:
|
||||
if not block_ids:
|
||||
return
|
||||
block_ids = np.array(block_ids)
|
||||
if self.use_hybrid_blocks:
|
||||
block_ids = self._convert_physical_to_logical_blocks(block_ids)
|
||||
|
||||
num_blocks = len(block_ids)
|
||||
start = self.num_blocks_per_row[row_idx]
|
||||
|
||||
self.block_table_np[row_idx, start:start + num_blocks] = block_ids
|
||||
self.num_blocks_per_row[row_idx] += num_blocks
|
||||
|
||||
def add_row(self, block_ids: list[int], row_idx: int) -> None:
|
||||
self.num_blocks_per_row[row_idx] = 0
|
||||
self.append_row(block_ids, row_idx)
|
||||
|
||||
def move_row(self, src: int, tgt: int) -> None:
|
||||
num_blocks = self.num_blocks_per_row[src]
|
||||
self.block_table_np[tgt, :num_blocks] = self.block_table_np[
|
||||
src, :num_blocks]
|
||||
self.num_blocks_per_row[tgt] = num_blocks
|
||||
|
||||
def swap_row(self, src: int, tgt: int) -> None:
|
||||
num_blocks_src = self.num_blocks_per_row[src]
|
||||
num_blocks_tgt = self.num_blocks_per_row[tgt]
|
||||
self.num_blocks_per_row[src] = num_blocks_tgt
|
||||
self.num_blocks_per_row[tgt] = num_blocks_src
|
||||
|
||||
self.block_table_np[[src, tgt]] = self.block_table_np[[tgt, src]]
|
||||
|
||||
def compute_slot_mapping(self, req_indices: np.ndarray,
|
||||
positions: np.ndarray) -> None:
|
||||
# E.g., [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
|
||||
# -> [0, 0, K, K, K + 1, K + 1, K + 2, 2 * K, 2 * K, 2 * K + 1]
|
||||
# where K is the max_num_blocks_per_req and the block size is 2.
|
||||
# NOTE(woosuk): We can't simply use `token_indices // block_size`
|
||||
# here because M (max_model_len) is not necessarily divisible by
|
||||
# block_size.
|
||||
|
||||
if self.dcp_world_size > 1:
|
||||
# Note(hc): The DCP implement store kvcache with an interleave
|
||||
# style, the kvcache for the token whose token_idx is i is
|
||||
# always stored on the GPU whose dcp_rank equals i % cp_world_size:
|
||||
|
||||
# Use a "virtual block" which equals to world_size * block_size
|
||||
# for block_table_indices calculation.
|
||||
virtual_block_size = self.block_size * self.dcp_world_size
|
||||
|
||||
# IMPORTANT: In hybrid mode, positions are in logical block space,
|
||||
# but we need to map them to the correct logical block table indices
|
||||
logical_block_idx = positions // virtual_block_size
|
||||
|
||||
# Account for the expanded logical table
|
||||
# (always needed with unified tensor)
|
||||
# Each physical block is split into multiple logical blocks
|
||||
# The logical table has been expanded to accommodate this
|
||||
block_table_indices = (req_indices * self.max_num_blocks_per_req *
|
||||
self.blocks_per_phys_block +
|
||||
logical_block_idx)
|
||||
|
||||
block_numbers = self.block_table_np.ravel()[block_table_indices]
|
||||
# Use virtual_block_size for mask calculation, which marks local
|
||||
# tokens.
|
||||
virtual_block_offsets = positions % virtual_block_size
|
||||
mask = virtual_block_offsets % self.dcp_world_size == self.dcp_rank
|
||||
# Calculate local block_offsets
|
||||
block_offsets = virtual_block_offsets // self.dcp_world_size
|
||||
# Calculate slot_mapping
|
||||
slot_mapping = block_numbers * self.block_size + block_offsets
|
||||
# Write final slots, use -1 for not-local
|
||||
self.slot_mapping_np[:req_indices.shape[0]] = np.where(
|
||||
mask, slot_mapping, -1)
|
||||
else:
|
||||
assert self.kernel_sizes is not None
|
||||
if self.block_size == self.kernel_sizes[0]:
|
||||
# IMPORTANT: In hybrid mode, positions are in logical block space,
|
||||
# but we need to map them to the correct logical block table indices
|
||||
logical_block_idx = positions // self.block_size
|
||||
|
||||
# Account for the expanded logical table
|
||||
# (always needed with unified tensor)
|
||||
# Each physical block is split into multiple logical blocks
|
||||
# The logical table has been expanded to accommodate this
|
||||
block_table_indices = (
|
||||
req_indices * self.max_num_blocks_per_req *
|
||||
self.blocks_per_phys_block + logical_block_idx)
|
||||
|
||||
block_numbers = self.block_table_np.ravel(
|
||||
)[block_table_indices]
|
||||
block_offsets = positions % self.block_size
|
||||
np.add(block_numbers * self.block_size,
|
||||
block_offsets,
|
||||
out=self.slot_mapping_np[:req_indices.shape[0]])
|
||||
|
||||
def commit_block_table(self, num_reqs: int) -> None:
|
||||
self.block_table[:num_reqs].copy_(self.block_table_cpu[:num_reqs],
|
||||
non_blocking=True)
|
||||
|
||||
def commit_slot_mapping(self, num_tokens: int) -> None:
|
||||
self.slot_mapping[:num_tokens].copy_(
|
||||
self.slot_mapping_cpu[:num_tokens], non_blocking=True)
|
||||
|
||||
def clear(self) -> None:
|
||||
self.block_table.fill_(0)
|
||||
self.block_table_cpu.fill_(0)
|
||||
|
||||
def _convert_physical_to_logical_blocks(
|
||||
self, physical_blocks: np.ndarray) -> np.ndarray:
|
||||
"""Convert physical block IDs to logical block IDs."""
|
||||
if not self.use_hybrid_blocks:
|
||||
return physical_blocks
|
||||
|
||||
# Create logical block IDs by splitting each physical block
|
||||
logical_blocks: list[int] = []
|
||||
for phys_block in physical_blocks:
|
||||
# Convert physical block to multiple logical blocks
|
||||
# Physical block 1 becomes logical blocks
|
||||
# [1*split_ratio, 1*split_ratio+1, ...]
|
||||
# But we need to account for the fact that block 0 is special
|
||||
base_logical = phys_block * self.blocks_per_phys_block
|
||||
logical_blocks.extend(
|
||||
range(base_logical, base_logical + self.blocks_per_phys_block))
|
||||
|
||||
return np.array(logical_blocks, dtype=np.int32)
|
||||
|
||||
def get_device_tensor(self) -> torch.Tensor:
|
||||
"""Returns the device tensor of the block table."""
|
||||
return self.block_table
|
||||
|
||||
def get_cpu_tensor(self) -> torch.Tensor:
|
||||
"""Returns the CPU tensor of the block table."""
|
||||
return self.block_table_cpu
|
||||
|
||||
def get_numpy_array(self) -> np.ndarray:
|
||||
"""Returns the numpy array of the block table."""
|
||||
return self.block_table_np
|
||||
|
||||
|
||||
class MultiGroupBlockTable:
|
||||
"""The BlockTables for each KV cache group."""
|
||||
|
||||
def __init__(self,
|
||||
max_num_reqs: int,
|
||||
max_model_len: int,
|
||||
max_num_batched_tokens: int,
|
||||
pin_memory: bool,
|
||||
device: torch.device,
|
||||
block_sizes: list[int],
|
||||
num_speculative_tokens: int = 0,
|
||||
kernel_sizes: Optional[list[list[int]]] = None) -> None:
|
||||
# Note(hc): each dcp rank only store
|
||||
# (max_model_len//dcp_world_size) tokens in kvcache,
|
||||
# so the block_size which used for calc max_num_blocks_per_req
|
||||
# must be multiplied by dcp_world_size.
|
||||
try:
|
||||
dcp_world_size = get_dcp_group().world_size
|
||||
except AssertionError:
|
||||
# DCP might not be initialized in testing
|
||||
dcp_world_size = 1
|
||||
|
||||
if kernel_sizes is None:
|
||||
kernel_sizes = [[0]] * len(block_sizes)
|
||||
# Ensure kernel_sizes matches block_sizes length
|
||||
elif len(kernel_sizes) == 1 and len(block_sizes) > 1:
|
||||
kernel_sizes = kernel_sizes * len(block_sizes)
|
||||
elif len(kernel_sizes) != len(block_sizes):
|
||||
raise ValueError(
|
||||
f"kernel_sizes length ({len(kernel_sizes)}) must match "
|
||||
f"block_sizes length ({len(block_sizes)})")
|
||||
|
||||
# Use zip to pair block_sizes with kernel_sizes one-to-one
|
||||
self.block_tables = [
|
||||
BlockTable(
|
||||
block_size, max_num_reqs,
|
||||
max(cdiv(max_model_len, block_size * dcp_world_size),
|
||||
1 + num_speculative_tokens), max_num_batched_tokens,
|
||||
pin_memory, device, kernel_size_list)
|
||||
for block_size, kernel_size_list in zip(block_sizes, kernel_sizes)
|
||||
]
|
||||
|
||||
def append_row(self, block_ids: tuple[list[int], ...],
|
||||
row_idx: int) -> None:
|
||||
for i, block_table in enumerate(self.block_tables):
|
||||
block_table.append_row(block_ids[i], row_idx)
|
||||
|
||||
def add_row(self, block_ids: tuple[list[int], ...], row_idx: int) -> None:
|
||||
for i, block_table in enumerate(self.block_tables):
|
||||
block_table.add_row(block_ids[i], row_idx)
|
||||
|
||||
def move_row(self, src: int, tgt: int) -> None:
|
||||
for block_table in self.block_tables:
|
||||
block_table.move_row(src, tgt)
|
||||
|
||||
def swap_row(self, src: int, tgt: int) -> None:
|
||||
for block_table in self.block_tables:
|
||||
block_table.swap_row(src, tgt)
|
||||
|
||||
def compute_slot_mapping(self, req_indices: np.ndarray,
|
||||
positions: np.ndarray) -> None:
|
||||
for block_table in self.block_tables:
|
||||
block_table.compute_slot_mapping(req_indices, positions)
|
||||
|
||||
def commit_block_table(self, num_reqs: int) -> None:
|
||||
for block_table in self.block_tables:
|
||||
block_table.commit_block_table(num_reqs)
|
||||
|
||||
def commit_slot_mapping(self, num_tokens: int) -> None:
|
||||
for block_table in self.block_tables:
|
||||
block_table.commit_slot_mapping(num_tokens)
|
||||
|
||||
def clear(self) -> None:
|
||||
for block_table in self.block_tables:
|
||||
block_table.clear()
|
||||
|
||||
def __getitem__(self, idx: int) -> "BlockTable":
|
||||
"""Returns the BlockTable for the i-th KV cache group."""
|
||||
return self.block_tables[idx]
|
||||
@@ -1,398 +0,0 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
import os
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from vllm.attention.layer import Attention
|
||||
from vllm.config import (CompilationLevel, VllmConfig,
|
||||
get_layers_from_vllm_config)
|
||||
from vllm.distributed.parallel_state import get_pp_group
|
||||
from vllm.logger import logger
|
||||
from vllm.model_executor.model_loader import get_model
|
||||
from vllm.model_executor.models import supports_multimodal
|
||||
from vllm.model_executor.models.llama_eagle3 import Eagle3LlamaForCausalLM
|
||||
from vllm.v1.sample.metadata import SamplingMetadata
|
||||
|
||||
from vllm_ascend.ascend_forward_context import set_ascend_forward_context
|
||||
from vllm_ascend.attention.attention_mask import AttentionMaskBuilder
|
||||
from vllm_ascend.attention.attention_v1 import AscendAttentionState
|
||||
from vllm_ascend.attention.utils import AscendCommonAttentionMetadata
|
||||
|
||||
PADDING_SLOT_ID = -1
|
||||
|
||||
|
||||
class EagleProposer:
|
||||
|
||||
def __init__(self,
|
||||
vllm_config: VllmConfig,
|
||||
device: torch.device,
|
||||
runner=None):
|
||||
self.vllm_config = vllm_config
|
||||
self.speculative_config = vllm_config.speculative_config
|
||||
self.draft_model_config = self.speculative_config.draft_model_config
|
||||
self.method = self.speculative_config.method
|
||||
self.runner = runner
|
||||
self.model_config = vllm_config.model_config
|
||||
self.dtype = vllm_config.model_config.dtype
|
||||
self.max_model_len = vllm_config.model_config.max_model_len
|
||||
self.block_size = vllm_config.cache_config.block_size
|
||||
self.num_speculative_tokens = (
|
||||
self.speculative_config.num_speculative_tokens)
|
||||
self.max_num_tokens = (
|
||||
vllm_config.scheduler_config.max_num_batched_tokens)
|
||||
self.device = device
|
||||
# We need to get the hidden size from the draft model config because
|
||||
# the draft model's hidden size can be different from the target model's
|
||||
# hidden size (e.g., Llama 3.3 70B).
|
||||
self.hidden_size = self.draft_model_config.get_hidden_size()
|
||||
|
||||
self.use_cuda_graph = (self.vllm_config.compilation_config.level
|
||||
== CompilationLevel.PIECEWISE and
|
||||
not self.vllm_config.model_config.enforce_eager)
|
||||
self.cudagraph_batch_sizes = list(
|
||||
reversed(
|
||||
self.vllm_config.compilation_config.cudagraph_capture_sizes))
|
||||
|
||||
# persistent buffers for cuda graph
|
||||
self.input_ids = torch.zeros(self.max_num_tokens,
|
||||
dtype=torch.int32,
|
||||
device=device)
|
||||
self.positions = torch.zeros(self.max_num_tokens,
|
||||
dtype=torch.int64,
|
||||
device=device)
|
||||
self.hidden_states = torch.zeros(
|
||||
(self.max_num_tokens, self.hidden_size),
|
||||
dtype=self.dtype,
|
||||
device=device)
|
||||
# We need +1 here because the arange is used to set query_start_loc,
|
||||
# which has one more element than batch_size.
|
||||
self.arange = torch.arange(vllm_config.scheduler_config.max_num_seqs +
|
||||
1,
|
||||
device=device,
|
||||
dtype=torch.int32)
|
||||
mask_len = os.getenv("PAGED_ATTENTION_MASK_LEN", 10000)
|
||||
self.attn_mask_len = min(self.model_config.max_model_len,
|
||||
int(mask_len))
|
||||
self.attn_mask_builder = AttentionMaskBuilder(self.attn_mask_len,
|
||||
self.dtype)
|
||||
|
||||
def _make_attention_mask(
|
||||
self,
|
||||
seq_lens,
|
||||
position,
|
||||
) -> torch.Tensor:
|
||||
return self.attn_mask_builder.get_splitfuse_attn_mask(
|
||||
seq_lens, position, self.dtype, self.device)
|
||||
|
||||
def propose(
|
||||
self,
|
||||
# [num_tokens]
|
||||
target_token_ids: torch.Tensor,
|
||||
# [num_tokens]
|
||||
target_positions: torch.Tensor,
|
||||
# [num_tokens, hidden_size]
|
||||
target_hidden_states: torch.Tensor,
|
||||
# [num_tokens]
|
||||
target_slot_mapping: torch.Tensor,
|
||||
# [batch_size]
|
||||
next_token_ids: torch.Tensor,
|
||||
# [batch_size + 1] starting with 0
|
||||
cu_num_tokens: torch.Tensor,
|
||||
# [batch_size, max_num_blocks_per_req]
|
||||
block_table: torch.Tensor,
|
||||
sampling_metadata: SamplingMetadata,
|
||||
) -> torch.Tensor:
|
||||
device = cu_num_tokens.device
|
||||
cu_num_tokens = cu_num_tokens.cpu()
|
||||
block_table = block_table.cpu()
|
||||
num_tokens = target_token_ids.shape[0]
|
||||
batch_size = next_token_ids.shape[0]
|
||||
last_token_indices = cu_num_tokens[1:] - 1
|
||||
target_positions = target_positions.cpu()
|
||||
if self.method == "eagle3":
|
||||
assert isinstance(self.model, Eagle3LlamaForCausalLM)
|
||||
target_hidden_states = self.model.combine_hidden_states(
|
||||
target_hidden_states)
|
||||
assert target_hidden_states.shape[-1] == self.hidden_size
|
||||
|
||||
# Shift the input ids by one token.
|
||||
# E.g., [a1, b1, b2, c1, c2, c3] -> [b1, b2, c1, c2, c3, c3]
|
||||
self.input_ids[:num_tokens - 1] = target_token_ids[1:]
|
||||
# Replace the last token with the next token.
|
||||
# E.g., [b1, b2, c1, c2, c3, c3] -> [a2, b2, b3, c2, c3, c4]
|
||||
self.input_ids[last_token_indices] = next_token_ids[0]
|
||||
|
||||
query_lens = cu_num_tokens[1:] - cu_num_tokens[:-1]
|
||||
max_query_len = query_lens.max().item()
|
||||
|
||||
common_attn_metadata = AscendCommonAttentionMetadata(
|
||||
query_start_loc=self.runner.query_start_loc[:batch_size + 1],
|
||||
query_start_loc_cpu=self.runner.query_start_loc_cpu[:batch_size +
|
||||
1],
|
||||
seq_lens_cpu=self.runner.seq_lens_cpu,
|
||||
max_query_len=max_query_len,
|
||||
num_reqs=batch_size,
|
||||
num_actual_tokens=num_tokens,
|
||||
actual_seq_lengths_q=self.runner.actual_seq_lengths_q,
|
||||
block_table_tensor=self.runner.input_batch.block_table[0].
|
||||
get_device_tensor(),
|
||||
slot_mapping_cpu=target_slot_mapping,
|
||||
positions=target_positions,
|
||||
attn_mask=self.runner.attn_mask,
|
||||
spec_attn_mask=self.runner.spec_attn_mask,
|
||||
attn_state=self.runner.attn_state,
|
||||
decode_token_per_req=self.runner.decode_token_per_req,
|
||||
)
|
||||
# FIXME(woosuk): The below two ops cause synchronization. Optimize.
|
||||
attn_metadata = self.runner.attn_metadata_builder.build(
|
||||
common_attn_metadata, self.runner.model)
|
||||
if self.use_cuda_graph and \
|
||||
num_tokens <= self.cudagraph_batch_sizes[-1]:
|
||||
num_input_tokens = self.vllm_config.pad_for_cudagraph(num_tokens)
|
||||
else:
|
||||
num_input_tokens = num_tokens
|
||||
# copy inputs to buffer for cudagraph
|
||||
self.positions[:num_tokens] = target_positions.to(device)
|
||||
self.hidden_states[:num_tokens] = target_hidden_states
|
||||
attn_metadata.block_tables = block_table.to(device)
|
||||
with set_ascend_forward_context(attn_metadata,
|
||||
self.vllm_config,
|
||||
num_tokens=num_input_tokens):
|
||||
last_hidden_states, hidden_states = self.model(
|
||||
input_ids=self.input_ids[:num_input_tokens],
|
||||
positions=self.positions[:num_input_tokens],
|
||||
hidden_states=self.hidden_states[:num_input_tokens],
|
||||
)
|
||||
sample_hidden_states = last_hidden_states[last_token_indices]
|
||||
logits = self.model.compute_logits(sample_hidden_states, None)
|
||||
draft_token_ids = logits.argmax(dim=-1)
|
||||
|
||||
# Early exit if there is only one draft token to be generated.
|
||||
if self.num_speculative_tokens == 1:
|
||||
# [batch_size, 1]
|
||||
return draft_token_ids.view(-1, 1)
|
||||
|
||||
# Generate the remaining draft tokens.
|
||||
draft_token_ids_tensor = torch.zeros(
|
||||
(self.num_speculative_tokens, *draft_token_ids.shape),
|
||||
dtype=draft_token_ids.dtype)
|
||||
draft_token_ids_tensor[0] = draft_token_ids
|
||||
|
||||
positions_cpu = target_positions[last_token_indices].cpu().to(
|
||||
torch.int64)
|
||||
hidden_states = hidden_states[last_token_indices]
|
||||
if self.use_cuda_graph and \
|
||||
batch_size <= self.cudagraph_batch_sizes[-1]:
|
||||
input_batch_size = self.vllm_config.pad_for_cudagraph(batch_size)
|
||||
else:
|
||||
input_batch_size = batch_size
|
||||
attn_metadata.num_actual_tokens = batch_size
|
||||
attn_metadata.max_query_len = 1
|
||||
attn_metadata.query_start_loc = self.arange[:batch_size + 1]
|
||||
|
||||
if self.num_speculative_tokens > 2:
|
||||
raise ValueError("Speculative tokens > 2 are not supported yet.")
|
||||
|
||||
attn_metadata.attn_state = AscendAttentionState.ChunkedPrefill
|
||||
for now_speculative in range(self.num_speculative_tokens - 1):
|
||||
# Update the inputs.
|
||||
# cast to int32 is crucial when eagle model is compiled.
|
||||
# tensor.argmax() returns int64 by default.
|
||||
input_ids = draft_token_ids_tensor[now_speculative].to(device)
|
||||
positions_cpu += 1
|
||||
|
||||
# NOTE(woosuk): We should handle the case where the draft model
|
||||
# generates tokens beyond the max model length. Since it is complex
|
||||
# to remove such requests from the batch, we keep them in the batch
|
||||
# but adjust the position ids and slot mappings to avoid the
|
||||
# out-of-range access during the model execution. The draft tokens
|
||||
# generated with this adjustment should be ignored.
|
||||
exceeds_max_model_len = positions_cpu >= self.max_model_len
|
||||
# Mask out the position ids that exceed the max model length.
|
||||
# Otherwise, we may get out-of-range error in RoPE.
|
||||
clamped_positions_cpu = torch.where(exceeds_max_model_len, 0,
|
||||
positions_cpu)
|
||||
clamped_positions = clamped_positions_cpu.to(device)
|
||||
|
||||
# TODO: Increment the sequence lengths.
|
||||
|
||||
attn_metadata.seq_lens += 1
|
||||
# TODO: Consider max model length.
|
||||
# attn_metadata.max_seq_len = min(attn_metadata.max_seq_len,
|
||||
# self.max_model_len)
|
||||
# For the requests that exceed the max model length, we set the
|
||||
# TODO: sequence length to 1 to minimize their overheads in attention.
|
||||
|
||||
# Compute the slot mapping.
|
||||
block_numbers = (clamped_positions_cpu // self.block_size)
|
||||
block_ids = block_table.gather(dim=1,
|
||||
index=block_numbers.view(-1, 1))
|
||||
block_ids = block_ids.view(-1)
|
||||
slot_mapping_cpu = (block_ids * self.block_size +
|
||||
clamped_positions_cpu % self.block_size)
|
||||
|
||||
# Mask out the slot mappings that exceed the max model length.
|
||||
# Otherwise, the KV cache will be inadvertently updated with the
|
||||
# padding tokens.
|
||||
slot_mapping_cpu.masked_fill_(exceeds_max_model_len,
|
||||
PADDING_SLOT_ID)
|
||||
# NOTE: ASCEND slot_mapping must on cpu
|
||||
attn_metadata.slot_mapping = slot_mapping_cpu.to(
|
||||
torch.int32).to(device)
|
||||
# copy inputs to buffer for cudagraph
|
||||
self.input_ids[:batch_size] = input_ids
|
||||
self.positions[:batch_size] = clamped_positions
|
||||
self.hidden_states[:batch_size] = hidden_states
|
||||
positions = positions_cpu.to(device)
|
||||
attn_mask = self._make_attention_mask(
|
||||
seq_lens=attn_metadata.seq_lens,
|
||||
position=positions,
|
||||
)
|
||||
attn_metadata.attn_mask = attn_mask
|
||||
attn_metadata.block_tables = block_table.to(device)
|
||||
# Run the model.
|
||||
with set_ascend_forward_context(attn_metadata,
|
||||
self.vllm_config,
|
||||
num_tokens=input_batch_size):
|
||||
|
||||
last_hidden_states, hidden_states = self.model(
|
||||
input_ids=self.input_ids[:input_batch_size],
|
||||
positions=self.positions[:input_batch_size],
|
||||
hidden_states=self.hidden_states[:input_batch_size],
|
||||
)
|
||||
hidden_states = hidden_states[:batch_size]
|
||||
logits = self.model.compute_logits(last_hidden_states[:batch_size],
|
||||
None)
|
||||
|
||||
# TODO(wenlong): get more than one token for tree attention
|
||||
draft_token_ids = logits.argmax(dim=-1)
|
||||
draft_token_ids_tensor[now_speculative + 1] = draft_token_ids.cpu()
|
||||
|
||||
# [batch_size, num_speculative_tokens]
|
||||
draft_token_ids = draft_token_ids_tensor.swapaxes(0, 1)
|
||||
return draft_token_ids
|
||||
|
||||
@staticmethod
|
||||
def prepare_inputs(
|
||||
# [batch_size + 1]
|
||||
cu_target_query_lens: torch.Tensor,
|
||||
# [batch_size]
|
||||
num_rejected_tokens: torch.Tensor,
|
||||
num_tokens: int,
|
||||
) -> tuple[torch.Tensor, torch.Tensor]:
|
||||
# cu_target_query_lens: [0, a, a + b, a + b + c]
|
||||
# num_rejected_tokens: [n1, n2, n3]
|
||||
# num_tokens_per_req: [a - n1, b - n2, c - n3]
|
||||
# cu_num_tokens: [0, a - n1, a + b - n1 - n2, a + b + c - n1 - n2 - n3]
|
||||
# token_indices: [0, 1, ..., a - n1 - 1,
|
||||
# a, a + 1, ..., a + b - n2 - 1,
|
||||
# a + b, a + b + 1, ..., a + b + c - n3 - 1]
|
||||
|
||||
# [0, a, a + b, a + b + c] -> [a, b, c]
|
||||
query_len_per_req = (cu_target_query_lens[1:] -
|
||||
cu_target_query_lens[:-1])
|
||||
# [a, b, c] -> [a - n1, b - n2, c - n3]
|
||||
num_tokens_per_req = query_len_per_req - num_rejected_tokens
|
||||
|
||||
# [a - n1, b - n2, c - n3] ->
|
||||
# [0, a - n1, a + b - n1 - n2, a + b + c - n1 - n2 - n3]
|
||||
cu_num_tokens = torch.zeros_like(cu_target_query_lens)
|
||||
torch.cumsum(num_tokens_per_req, dim=0, out=cu_num_tokens[1:])
|
||||
token_indices = torch.empty(
|
||||
num_tokens,
|
||||
dtype=torch.int32,
|
||||
device=cu_target_query_lens.device,
|
||||
)
|
||||
BLOCK_SIZE = 1024
|
||||
prepare_eagle_input_sequential(
|
||||
token_indices,
|
||||
cu_target_query_lens,
|
||||
cu_num_tokens,
|
||||
block_size=BLOCK_SIZE,
|
||||
)
|
||||
return cu_num_tokens, token_indices
|
||||
|
||||
def load_model(self, target_model: nn.Module) -> None:
|
||||
draft_model_config = \
|
||||
self.vllm_config.speculative_config.draft_model_config
|
||||
target_attn_layer_names = set(
|
||||
get_layers_from_vllm_config(self.vllm_config, Attention).keys())
|
||||
|
||||
self.model = get_model(vllm_config=self.vllm_config,
|
||||
model_config=draft_model_config)
|
||||
|
||||
draft_attn_layer_names = (
|
||||
get_layers_from_vllm_config(self.vllm_config, Attention).keys() -
|
||||
target_attn_layer_names)
|
||||
|
||||
self.attn_layer_names = list(draft_attn_layer_names)
|
||||
self.attn_layer_name = next(iter(draft_attn_layer_names))
|
||||
# share embed_tokens with the target model if needed
|
||||
if get_pp_group().world_size == 1:
|
||||
logger.info(
|
||||
"The EAGLE head shares the same vocab embedding" \
|
||||
" with the target model."
|
||||
)
|
||||
self.model.model.embed_tokens = target_model.model.embed_tokens
|
||||
else:
|
||||
logger.info(
|
||||
"Since PP > 1, the EAGLE head loaded its own vocab embedding" \
|
||||
" weights instead of sharing them with the target model."
|
||||
)
|
||||
|
||||
# share lm_head with the target model if needed
|
||||
# some model definition do not define lm_head explicitly
|
||||
# and reuse embed_tokens for lm_head, e.g., CohereForCausalLM
|
||||
if self.vllm_config.speculative_config.method != "eagle3" and \
|
||||
hasattr(target_model, "lm_head"):
|
||||
logger.info("Loading EAGLE LM head weights from the target model.")
|
||||
if supports_multimodal(target_model):
|
||||
self.model.lm_head = target_model.get_language_model().lm_head
|
||||
else:
|
||||
self.model.lm_head = target_model.lm_head
|
||||
|
||||
@torch.inference_mode()
|
||||
def dummy_run(
|
||||
self,
|
||||
num_tokens: int,
|
||||
) -> None:
|
||||
with set_ascend_forward_context(None,
|
||||
self.vllm_config,
|
||||
num_tokens=num_tokens):
|
||||
self.model(
|
||||
input_ids=self.input_ids[:num_tokens],
|
||||
positions=self.positions[:num_tokens],
|
||||
hidden_states=self.hidden_states[:num_tokens],
|
||||
)
|
||||
|
||||
|
||||
def prepare_eagle_input_sequential(out_tensor: torch.Tensor,
|
||||
cu_query_lens: torch.Tensor,
|
||||
cu_num_tokens: torch.Tensor,
|
||||
block_size: int):
|
||||
num_programs = len(cu_num_tokens) - 1
|
||||
for pid in range(num_programs):
|
||||
start_pos = cu_num_tokens[pid].item()
|
||||
end_pos = cu_num_tokens[pid + 1].item()
|
||||
num_tokens = end_pos - start_pos
|
||||
index_start = cu_query_lens[pid].item()
|
||||
num_blocks = int(
|
||||
torch.ceil(torch.tensor(num_tokens / block_size)).item())
|
||||
|
||||
for i in range(num_blocks):
|
||||
offset_tensor = torch.arange(0,
|
||||
block_size,
|
||||
dtype=torch.int32,
|
||||
device=out_tensor.device)
|
||||
global_start_offset = i * block_size
|
||||
target_indices = torch.tensor(
|
||||
start_pos + global_start_offset,
|
||||
dtype=torch.int32,
|
||||
device=out_tensor.device) + offset_tensor
|
||||
values_to_store = torch.tensor(
|
||||
index_start, dtype=torch.int32,
|
||||
device=out_tensor.device) + offset_tensor
|
||||
mask = (target_indices >= start_pos) & \
|
||||
(target_indices < end_pos) & \
|
||||
(offset_tensor < num_tokens)
|
||||
out_tensor[target_indices[mask]] = values_to_store[mask]
|
||||
File diff suppressed because it is too large
Load Diff
@@ -1,439 +0,0 @@
|
||||
import types
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import torchair
|
||||
import vllm.envs as envs_vllm
|
||||
from torchair import patch_for_hcom
|
||||
from vllm.attention.layer import Attention
|
||||
from vllm.config import (VllmConfig, get_layers_from_vllm_config,
|
||||
set_current_vllm_config)
|
||||
from vllm.forward_context import get_forward_context
|
||||
from vllm.model_executor.model_loader import get_model_loader
|
||||
from vllm.model_executor.model_loader.utils import (
|
||||
process_weights_after_loading, set_default_torch_dtype)
|
||||
from vllm.v1.sample.metadata import SamplingMetadata
|
||||
|
||||
from vllm_ascend.ascend_config import get_ascend_config
|
||||
from vllm_ascend.ascend_forward_context import set_ascend_forward_context
|
||||
from vllm_ascend.attention.utils import AscendCommonAttentionMetadata
|
||||
from vllm_ascend.models.deepseek_mtp import CustomDeepSeekMTP
|
||||
from vllm_ascend.torchair.models.torchair_deepseek_mtp import \
|
||||
TorchairDeepSeekMTP
|
||||
from vllm_ascend.torchair.utils import (TORCHAIR_CACHE_DIR,
|
||||
TorchairCommonAttentionMetadata)
|
||||
from vllm_ascend.utils import ProfileExecuteDuration, lmhead_tp_enable
|
||||
|
||||
|
||||
class MtpProposer:
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
vllm_config: VllmConfig,
|
||||
runner,
|
||||
):
|
||||
self.vllm_config = vllm_config
|
||||
self.num_speculative_tokens = (
|
||||
vllm_config.speculative_config.num_speculative_tokens)
|
||||
self.block_size = vllm_config.cache_config.block_size
|
||||
self.hidden_size = vllm_config.model_config.get_hidden_size()
|
||||
self.runner = runner
|
||||
# persistent buffers for graph
|
||||
self.input_ids = torch.zeros(self.runner.max_num_tokens,
|
||||
dtype=torch.int32,
|
||||
device=self.runner.device)
|
||||
self.positions = torch.zeros(self.runner.max_num_tokens,
|
||||
dtype=torch.int64,
|
||||
device=self.runner.device)
|
||||
self.hidden_states = torch.zeros(
|
||||
(self.runner.max_num_tokens, self.hidden_size),
|
||||
dtype=self.runner.dtype,
|
||||
device=self.runner.device)
|
||||
self.torchair_compiled_model = None # type: ignore
|
||||
self.torchair_compiled_models = {} # type: ignore
|
||||
self.torchair_graph_enabled = get_ascend_config(
|
||||
).torchair_graph_config.enabled
|
||||
|
||||
@staticmethod
|
||||
def prepare_inputs(
|
||||
# [batch_size + 1]
|
||||
cu_target_query_lens: torch.Tensor,
|
||||
# [batch_size]
|
||||
num_rejected_tokens: torch.Tensor,
|
||||
token_ids: torch.Tensor,
|
||||
positions: torch.Tensor,
|
||||
hidden_states: torch.Tensor,
|
||||
slot_mapping: torch.Tensor,
|
||||
is_torchair_graph: bool = False
|
||||
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor,
|
||||
torch.Tensor, torch.Tensor]:
|
||||
# cu_target_query_lens: [0, a, a + b, a + b + c]
|
||||
# num_rejected_tokens: [n1, n2, n3]
|
||||
# num_tokens_per_req: [a - n1, b - n2, c - n3]
|
||||
# cu_num_tokens: [0, a - n1, a + b - n1 - n2, a + b + c - n1 - n2 - n3]
|
||||
# token_indices: [0, 1, ..., a - n1 - 1,
|
||||
# a, a + 1, ..., a + b - n2 - 1,
|
||||
# a + b, a + b + 1, ..., a + b + c - n3 - 1]
|
||||
# [0, a, a + b, a + b + c] -> [a, b, c]
|
||||
query_len_per_req = (cu_target_query_lens[1:] -
|
||||
cu_target_query_lens[:-1])
|
||||
# [a, b, c] -> [a - n1, b - n2, c - n3]
|
||||
num_tokens_per_req = query_len_per_req - num_rejected_tokens
|
||||
if is_torchair_graph:
|
||||
cu_num_tokens = cu_target_query_lens
|
||||
relative_index = query_len_per_req - num_rejected_tokens - 1
|
||||
token_indices = cu_num_tokens[:-1] + relative_index
|
||||
# the seq len of each bath is padded to 1+num_speculative_tokens, thus input is same as the main model
|
||||
target_token_ids = token_ids
|
||||
target_positions = positions
|
||||
target_hidden_states = hidden_states
|
||||
target_slot_mapping = slot_mapping
|
||||
else:
|
||||
cu_num_tokens = torch.empty_like(cu_target_query_lens)
|
||||
torch.cumsum(num_tokens_per_req, dim=0, out=cu_num_tokens[1:])
|
||||
cu_num_tokens[0] = 0
|
||||
|
||||
# FIXME(woosuk): Avoid synchronization.
|
||||
num_tokens = cu_num_tokens[-1].item()
|
||||
token_indices = torch.zeros(
|
||||
num_tokens,
|
||||
dtype=torch.int32,
|
||||
device=cu_num_tokens.device,
|
||||
)
|
||||
|
||||
BLOCK_SIZE = 1024
|
||||
prepare_input_kernel(
|
||||
token_indices,
|
||||
cu_target_query_lens,
|
||||
cu_num_tokens,
|
||||
block_size=BLOCK_SIZE,
|
||||
)
|
||||
target_token_ids = token_ids[token_indices]
|
||||
target_positions = positions[token_indices]
|
||||
target_hidden_states = hidden_states[token_indices]
|
||||
target_slot_mapping = slot_mapping[token_indices]
|
||||
return cu_num_tokens, token_indices, target_token_ids, target_positions, target_hidden_states, target_slot_mapping
|
||||
|
||||
def propose(
|
||||
self,
|
||||
# [num_tokens]
|
||||
target_token_ids: torch.Tensor,
|
||||
# [num_tokens]
|
||||
target_positions: torch.Tensor,
|
||||
# [num_tokens, hidden_size]
|
||||
target_hidden_states: torch.Tensor,
|
||||
# [num_tokens]
|
||||
target_slot_mapping: torch.Tensor,
|
||||
# [batch_size]
|
||||
next_token_ids: torch.Tensor,
|
||||
# [batch_size + 1] starting with 0
|
||||
cu_num_tokens: torch.Tensor,
|
||||
# [batch_size, max_num_blocks_per_req]
|
||||
block_table: torch.Tensor,
|
||||
sampling_metadata: SamplingMetadata,
|
||||
token_indices=None) -> torch.Tensor:
|
||||
num_tokens = target_token_ids.shape[0]
|
||||
batch_size = next_token_ids.shape[0]
|
||||
last_token_indices = cu_num_tokens[1:] - 1
|
||||
|
||||
# Shift the input ids by one token.
|
||||
# E.g., [a1, b1, b2, c1, c2, c3] -> [b1, b2, c1, c2, c3, c3]
|
||||
self.input_ids[:num_tokens - 1] = target_token_ids[1:]
|
||||
# Replace the last token with the next token.
|
||||
# E.g., [b1, b2, c1, c2, c3, c3] -> [a2, b2, b3, c2, c3, c4]
|
||||
if token_indices is not None and self.torchair_graph_enabled:
|
||||
last_token_indices = token_indices
|
||||
|
||||
self.input_ids[last_token_indices] = next_token_ids
|
||||
|
||||
query_lens = cu_num_tokens[1:] - cu_num_tokens[:-1]
|
||||
max_query_len = query_lens.max().item()
|
||||
|
||||
# FIXME: reorder_batch() needs to be called before build()
|
||||
# because fields of attn_metadata_builder needs to be updated.
|
||||
# However, currently reorder_batch() takes input_batch and
|
||||
# scheduler_output as arguments, we should probably refactor
|
||||
# the method to use new data structures which are independent
|
||||
# from input_batch and scheduler_output.
|
||||
# self.runner.attn_metadata_builder.reorder_batch(
|
||||
# input_batch=self.runner.input_batch,
|
||||
# scheduler_output=self.runner.scheduler_output,
|
||||
# )
|
||||
is_running_torchair = self.torchair_graph_enabled and \
|
||||
not self.runner.with_prefill
|
||||
|
||||
if is_running_torchair:
|
||||
num_input_tokens = self.runner.graph_pad_size
|
||||
else:
|
||||
num_input_tokens = num_tokens
|
||||
|
||||
seq_lens = target_positions[last_token_indices] + 1
|
||||
seq_lens = seq_lens.int()
|
||||
common_attn_metadata = AscendCommonAttentionMetadata(
|
||||
query_start_loc=cu_num_tokens[:batch_size + 1],
|
||||
query_start_loc_cpu=cu_num_tokens[:batch_size + 1].cpu(),
|
||||
seq_lens_cpu=seq_lens.cpu(),
|
||||
num_reqs=batch_size,
|
||||
num_actual_tokens=num_tokens,
|
||||
max_query_len=max_query_len,
|
||||
actual_seq_lengths_q=self.runner.actual_seq_lengths_q,
|
||||
block_table_tensor=self.runner.input_batch.block_table[0].
|
||||
get_device_tensor(),
|
||||
slot_mapping_cpu=target_slot_mapping,
|
||||
positions=target_positions,
|
||||
attn_mask=self.runner.attn_mask,
|
||||
spec_attn_mask=self.runner.spec_attn_mask,
|
||||
attn_state=self.runner.attn_state,
|
||||
graph_pad_size=self.runner.graph_pad_size,
|
||||
decode_token_per_req=self.runner.decode_token_per_req,
|
||||
)
|
||||
attn_metadata = self.runner.attn_metadata_builder.build(
|
||||
common_attn_metadata, self.runner.get_model())
|
||||
|
||||
self.positions[:num_tokens] = target_positions
|
||||
self.hidden_states[:num_tokens] = target_hidden_states
|
||||
|
||||
if not self.torchair_graph_enabled:
|
||||
# torch mode need to update num_tokens_across_dp
|
||||
# TODO: adapt enable_dbo later
|
||||
(num_input_tokens, num_tokens_across_dp, with_prefill,
|
||||
_) = self.runner._sync_metadata_across_dp(
|
||||
num_tokens, self.runner.with_prefill, False)
|
||||
attn_metadata.slot_mapping = target_slot_mapping
|
||||
else:
|
||||
# torchair mode can reuse self.runner.num_tokens_across_dp
|
||||
num_tokens_across_dp = self.runner.num_tokens_across_dp
|
||||
with_prefill = self.runner.with_prefill
|
||||
|
||||
with set_ascend_forward_context(
|
||||
attn_metadata,
|
||||
self.vllm_config,
|
||||
num_tokens=num_input_tokens,
|
||||
with_prefill=with_prefill,
|
||||
num_tokens_across_dp=num_tokens_across_dp,
|
||||
reserved_mc2_mask=self.runner.reserved_mc2_mask,
|
||||
in_profile_run=self.runner.in_profile_run,
|
||||
num_actual_tokens=num_tokens):
|
||||
with ProfileExecuteDuration().capture_async('mtp_forward'):
|
||||
model_kwargs = {}
|
||||
model_kwargs["attn_metadata"] = attn_metadata
|
||||
if self.torchair_graph_enabled:
|
||||
model_kwargs["kv_caches"] = self.runner.kv_caches[-1:]
|
||||
if is_running_torchair:
|
||||
torchair_compiled_model = self._get_torchair_lazy_compiled_model(
|
||||
num_input_tokens)
|
||||
hidden_states = torchair_compiled_model(
|
||||
input_ids=self.input_ids[:num_input_tokens],
|
||||
positions=self.positions[:num_input_tokens],
|
||||
previous_hidden_states=self.
|
||||
hidden_states[:num_input_tokens],
|
||||
inputs_embeds=None,
|
||||
intermediate_tensors=None,
|
||||
spec_step_idx=0,
|
||||
**model_kwargs)
|
||||
else:
|
||||
hidden_states = self.model(
|
||||
input_ids=self.input_ids[:num_input_tokens],
|
||||
positions=self.positions[:num_input_tokens],
|
||||
previous_hidden_states=self.
|
||||
hidden_states[:num_input_tokens],
|
||||
kv_caches=self.runner.kv_caches[-1:])
|
||||
|
||||
num_indices = last_token_indices.shape[0]
|
||||
if lmhead_tp_enable():
|
||||
if not self.runner.with_prefill:
|
||||
max_num_reqs_across_dp = num_input_tokens
|
||||
else:
|
||||
max_num_reqs_across_dp = self.vllm_config.scheduler_config.max_num_seqs
|
||||
last_token_indices = nn.functional.pad(
|
||||
last_token_indices, (0, max_num_reqs_across_dp - num_indices))
|
||||
|
||||
sample_hidden_states = hidden_states[last_token_indices]
|
||||
logits = self.model.compute_logits(sample_hidden_states, None)
|
||||
if lmhead_tp_enable() and num_indices < logits.shape[0]:
|
||||
logits = logits[:num_indices]
|
||||
draft_token_ids = logits.argmax(dim=-1)
|
||||
|
||||
# [batch_size, 1]
|
||||
return draft_token_ids.view(-1, 1)
|
||||
|
||||
def load_model(self) -> None:
|
||||
loader = get_model_loader(self.vllm_config.load_config)
|
||||
|
||||
target_attn_layer_names = set(
|
||||
get_layers_from_vllm_config(self.vllm_config, Attention).keys())
|
||||
draft_model_config = \
|
||||
self.vllm_config.speculative_config.draft_model_config
|
||||
target_device = self.vllm_config.device_config.device
|
||||
|
||||
with set_default_torch_dtype(
|
||||
draft_model_config.dtype), set_current_vllm_config(
|
||||
self.vllm_config):
|
||||
if self.torchair_graph_enabled:
|
||||
self.model = TorchairDeepSeekMTP(
|
||||
vllm_config=self.vllm_config).to(target_device)
|
||||
else:
|
||||
self.model = CustomDeepSeekMTP(
|
||||
vllm_config=self.vllm_config).to(target_device)
|
||||
|
||||
draft_attn_layer_names = (
|
||||
get_layers_from_vllm_config(self.vllm_config, Attention).keys() -
|
||||
target_attn_layer_names)
|
||||
|
||||
assert len(draft_attn_layer_names) == 1
|
||||
self.attn_layer_name = next(iter(draft_attn_layer_names))
|
||||
|
||||
self.model.load_weights(
|
||||
loader.get_all_weights(
|
||||
self.vllm_config.speculative_config.draft_model_config,
|
||||
self.model))
|
||||
process_weights_after_loading(self.model, draft_model_config,
|
||||
target_device)
|
||||
|
||||
@torch.inference_mode()
|
||||
def dummy_run(self,
|
||||
num_tokens: int,
|
||||
with_prefill: bool = False,
|
||||
skip_attn: bool = False,
|
||||
num_reqs: int = 0,
|
||||
num_tokens_across_dp=None) -> None:
|
||||
if not self.torchair_graph_enabled:
|
||||
# TODO: adapt enable_dbo later
|
||||
(num_tokens, num_tokens_across_dp, with_prefill,
|
||||
_) = self.runner._sync_metadata_across_dp(num_tokens,
|
||||
with_prefill, False)
|
||||
is_running_torchair = self.torchair_graph_enabled and \
|
||||
not with_prefill
|
||||
|
||||
if is_running_torchair:
|
||||
skip_attn = False
|
||||
if skip_attn:
|
||||
attn_metadata = None
|
||||
else:
|
||||
common_attn_metadata = TorchairCommonAttentionMetadata(
|
||||
num_reqs=num_reqs,
|
||||
num_actual_tokens=1,
|
||||
actual_seq_lengths_q=self.runner.actual_seq_lengths_q,
|
||||
attn_mask=self.runner.attn_mask,
|
||||
spec_attn_mask=self.runner.spec_attn_mask,
|
||||
decode_token_per_req=self.runner.decode_token_per_req,
|
||||
)
|
||||
attn_metadata = self.runner.attn_metadata_builder.build_torchair_graph_dummy(
|
||||
common_attn_metadata)
|
||||
|
||||
input_ids = self.input_ids[:num_tokens]
|
||||
positions = self.positions[:num_tokens]
|
||||
previous_hidden_states = self.hidden_states[:num_tokens]
|
||||
with set_ascend_forward_context(
|
||||
attn_metadata,
|
||||
self.vllm_config,
|
||||
num_tokens=num_tokens,
|
||||
with_prefill=with_prefill,
|
||||
num_tokens_across_dp=num_tokens_across_dp,
|
||||
reserved_mc2_mask=self.runner.reserved_mc2_mask,
|
||||
in_profile_run=self.runner.in_profile_run,
|
||||
num_actual_tokens=0):
|
||||
if is_running_torchair:
|
||||
assert attn_metadata is not None
|
||||
torch._dynamo.mark_static(input_ids)
|
||||
torch._dynamo.mark_static(positions)
|
||||
torch._dynamo.mark_static(previous_hidden_states)
|
||||
torch._dynamo.mark_static(attn_metadata.decode.block_table)
|
||||
torch._dynamo.mark_static(attn_metadata.decode.input_positions)
|
||||
if hasattr(attn_metadata.decode, "sin"):
|
||||
torch._dynamo.mark_static(attn_metadata.decode.sin)
|
||||
torch._dynamo.mark_static(attn_metadata.decode.cos)
|
||||
torch._dynamo.mark_static(get_forward_context().mc2_mask)
|
||||
torch._dynamo.mark_static(attn_metadata.slot_mapping)
|
||||
torch._dynamo.mark_static(attn_metadata.decode.attn_mask)
|
||||
torchair_compiled_model = self._get_torchair_lazy_compiled_model(
|
||||
num_tokens)
|
||||
torchair_compiled_model(
|
||||
input_ids=input_ids,
|
||||
positions=positions,
|
||||
previous_hidden_states=previous_hidden_states,
|
||||
inputs_embeds=None,
|
||||
intermediate_tensors=None,
|
||||
attn_metadata=attn_metadata,
|
||||
kv_caches=self.runner.kv_caches[-1:],
|
||||
spec_step_idx=0)
|
||||
else:
|
||||
self.model(input_ids=input_ids,
|
||||
positions=positions,
|
||||
previous_hidden_states=previous_hidden_states)
|
||||
|
||||
def _get_torchair_lazy_compiled_model(self, batch_size: int):
|
||||
if batch_size < 0 or batch_size > self.runner.torchair_graph_batch_sizes[
|
||||
-1]:
|
||||
raise ValueError(
|
||||
f"Bad graph batch size:{batch_size}! max_graph_batch_sizes:{self.runner.torchair_graph_batch_sizes[-1]}"
|
||||
)
|
||||
|
||||
compiled_model = self.torchair_compiled_models.get(
|
||||
batch_size
|
||||
) if self.runner.use_cached_npu_graph else self.torchair_compiled_model
|
||||
|
||||
if compiled_model:
|
||||
return compiled_model
|
||||
|
||||
patch_for_hcom()
|
||||
config = torchair.CompilerConfig()
|
||||
config.experimental_config.frozen_parameter = True
|
||||
config.experimental_config.tiling_schedule_optimize = True
|
||||
config.experimental_config.enable_view_optimize = \
|
||||
get_ascend_config().torchair_graph_config.enable_view_optimize
|
||||
torch.npu.set_compile_mode(jit_compile=False)
|
||||
if not self.runner.use_cached_npu_graph:
|
||||
npu_backend = torchair.get_npu_backend(compiler_config=config)
|
||||
self.torchair_compiled_model = torch.compile(
|
||||
self.model,
|
||||
dynamic=True,
|
||||
fullgraph=envs_vllm.VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE,
|
||||
backend=npu_backend)
|
||||
return self.torchair_compiled_model
|
||||
else:
|
||||
# Generate a new forward proxy code object to prevent the invalidation of
|
||||
# compilation cache caused by dynamo retracing
|
||||
forward_proxy_name = f"{self.model.__class__.__name__}_forward_with_batch_size_{batch_size}"
|
||||
forward_fn = self.model.forward
|
||||
code = forward_fn.__code__
|
||||
# Mark code object with a new proxy name
|
||||
modified_code = code.replace(co_name=forward_proxy_name, )
|
||||
|
||||
modified_func = types.FunctionType(modified_code,
|
||||
forward_fn.__globals__,
|
||||
name=forward_proxy_name,
|
||||
argdefs=forward_fn.__defaults__)
|
||||
|
||||
self.model.__dict__[forward_proxy_name] = modified_func.__get__(
|
||||
self.model, nn.Module)
|
||||
self.torchair_compiled_models[
|
||||
batch_size] = torchair.inference.cache_compile(
|
||||
self.model.__dict__[forward_proxy_name],
|
||||
dynamic=True,
|
||||
fullgraph=envs_vllm.VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE,
|
||||
cache_dir=TORCHAIR_CACHE_DIR,
|
||||
config=config,
|
||||
ge_cache=False)
|
||||
return self.torchair_compiled_models[batch_size]
|
||||
|
||||
|
||||
# TODO Using torch instead of triton may result in poor performance
|
||||
def prepare_input_kernel(out_ptr: torch.Tensor, cu_query_lens: torch.Tensor,
|
||||
cu_num_tokens: torch.Tensor, block_size: int):
|
||||
device = cu_query_lens.device
|
||||
dtype = out_ptr.dtype
|
||||
|
||||
offsets = torch.arange(block_size, device=device, dtype=dtype)
|
||||
start_pos = cu_num_tokens[:-1]
|
||||
end_pos = cu_num_tokens[1:]
|
||||
num_tokens = end_pos - start_pos
|
||||
|
||||
global_indices = (start_pos.view(-1, 1) + offsets.view(1, -1))
|
||||
values = (cu_query_lens[:-1].view(-1, 1) + offsets.view(1, -1))
|
||||
|
||||
mask = (offsets.view(1, -1) < num_tokens.view(-1, 1))
|
||||
|
||||
global_indices_flat = global_indices[mask]
|
||||
values_flat = values[mask]
|
||||
out_ptr[global_indices_flat] = values_flat
|
||||
@@ -24,8 +24,9 @@ import numpy as np
|
||||
import torch
|
||||
from typing_extensions import deprecated
|
||||
from vllm.lora.request import LoRARequest
|
||||
from vllm.multimodal.inputs import (MultiModalKwargs, MultiModalKwargsItem,
|
||||
PlaceholderRange)
|
||||
from vllm.multimodal.inputs import (MultiModalFeatureSpec,
|
||||
MultiModalKwargsItem,
|
||||
MultiModalKwargsItems, PlaceholderRange)
|
||||
from vllm.pooling_params import PoolingParams
|
||||
from vllm.sampling_params import SamplingParams, SamplingType
|
||||
from vllm.utils import swap_dict_values
|
||||
@@ -37,9 +38,9 @@ from vllm.v1.sample.logits_processor import (BatchUpdateBuilder,
|
||||
from vllm.v1.sample.metadata import SamplingMetadata
|
||||
from vllm.v1.spec_decode.utils import is_spec_decode_unsupported
|
||||
from vllm.v1.utils import copy_slice
|
||||
from vllm.v1.worker.block_table import MultiGroupBlockTable
|
||||
|
||||
from vllm_ascend.utils import vllm_version_is
|
||||
from vllm_ascend.worker.block_table import MultiGroupBlockTable
|
||||
|
||||
|
||||
@dataclass
|
||||
@@ -47,10 +48,6 @@ class CachedRequestState:
|
||||
|
||||
req_id: str
|
||||
prompt_token_ids: list[int]
|
||||
mm_kwargs: list[MultiModalKwargsItem]
|
||||
mm_positions: list[PlaceholderRange]
|
||||
# TODO: remove Optional after 0.10.1.1
|
||||
mm_hashes: Optional[list[str]]
|
||||
sampling_params: Optional[SamplingParams]
|
||||
pooling_params: Optional[PoolingParams]
|
||||
generator: Optional[torch.Generator]
|
||||
@@ -62,6 +59,12 @@ class CachedRequestState:
|
||||
mrope_positions: Optional[torch.Tensor] = None
|
||||
mrope_position_delta: Optional[int] = None
|
||||
|
||||
mm_features: Optional[list[MultiModalFeatureSpec]] = None
|
||||
# for back-compatibility, will be removed in next major release
|
||||
mm_kwargs: Optional[list[MultiModalKwargsItem]] = None
|
||||
mm_positions: Optional[list[PlaceholderRange]] = None
|
||||
mm_hashes: Optional[list[PlaceholderRange]] = None
|
||||
|
||||
lora_request: Optional[LoRARequest] = None
|
||||
|
||||
def __post_init__(self):
|
||||
@@ -75,8 +78,18 @@ class CachedRequestState:
|
||||
@property
|
||||
@deprecated("`mm_inputs` is superseded by `mm_kwargs` and will be "
|
||||
"removed in v0.13. Please use `mm_kwargs` instead.")
|
||||
def mm_inputs(self) -> list[MultiModalKwargs]:
|
||||
return [MultiModalKwargs([item]) for item in self.mm_kwargs]
|
||||
def mm_inputs(self) -> list[MultiModalKwargsItems]:
|
||||
if vllm_version_is("0.10.2"):
|
||||
assert self.mm_kwargs is not None
|
||||
return [
|
||||
MultiModalKwargsItems.from_seq([item])
|
||||
for item in self.mm_kwargs
|
||||
]
|
||||
assert self.mm_features is not None
|
||||
return [
|
||||
MultiModalKwargsItems.from_seq([f.data]) for f in self.mm_features
|
||||
if f.data is not None
|
||||
]
|
||||
|
||||
def get_token_id(self, idx: int) -> int:
|
||||
if idx < self.num_prompt_tokens:
|
||||
@@ -88,18 +101,19 @@ class CachedRequestState:
|
||||
class InputBatch:
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
max_num_reqs: int,
|
||||
max_model_len: int,
|
||||
max_num_batched_tokens: int,
|
||||
device: torch.device,
|
||||
pin_memory: bool,
|
||||
vocab_size: int,
|
||||
block_sizes: list[int], # The block_size of each kv cache group
|
||||
logitsprocs: Optional[LogitsProcessors] = None,
|
||||
is_spec_decode: bool = False,
|
||||
is_pooling_model: bool = False,
|
||||
):
|
||||
self,
|
||||
max_num_reqs: int,
|
||||
max_model_len: int,
|
||||
max_num_batched_tokens: int,
|
||||
device: torch.device,
|
||||
pin_memory: bool,
|
||||
vocab_size: int,
|
||||
block_sizes: list[int], # The block_size of each kv cache group
|
||||
logitsprocs: Optional[LogitsProcessors] = None,
|
||||
is_spec_decode: bool = False,
|
||||
is_pooling_model: bool = False,
|
||||
num_speculative_tokens: int = 0,
|
||||
kernel_block_sizes: Optional[list[list[int]]] = None):
|
||||
self.is_pooling_model = is_pooling_model
|
||||
self.is_spec_decode = is_spec_decode
|
||||
self.max_num_reqs = max_num_reqs
|
||||
@@ -143,7 +157,8 @@ class InputBatch:
|
||||
pin_memory=pin_memory,
|
||||
device=device,
|
||||
block_sizes=block_sizes,
|
||||
)
|
||||
num_speculative_tokens=num_speculative_tokens,
|
||||
kernel_sizes=kernel_block_sizes)
|
||||
|
||||
# Sampling-related.
|
||||
self.temperature = torch.empty((max_num_reqs, ),
|
||||
@@ -218,6 +233,14 @@ class InputBatch:
|
||||
self.repetition_penalties_cpu_tensor.numpy()
|
||||
self.repetition_penalties_reqs: set[str] = set()
|
||||
|
||||
# Speculative decoding
|
||||
self.num_accepted_tokens_cpu_tensor = torch.ones((max_num_reqs, ),
|
||||
dtype=torch.int64,
|
||||
device="cpu",
|
||||
pin_memory=pin_memory)
|
||||
self.num_accepted_tokens_cpu = \
|
||||
self.num_accepted_tokens_cpu_tensor.numpy()
|
||||
|
||||
# lora related
|
||||
self.request_lora_mapping = np.zeros((self.max_num_reqs, ),
|
||||
dtype=np.int32)
|
||||
@@ -266,6 +289,11 @@ class InputBatch:
|
||||
|
||||
self.pooling_params: dict[str, PoolingParams] = {}
|
||||
|
||||
# Cached reference to the GPU tensor of previously sampled tokens
|
||||
self.prev_sampled_token_ids: Optional[torch.Tensor] = None
|
||||
self.prev_sampled_token_ids_invalid_indices: Optional[set[int]] = None
|
||||
self.prev_req_id_to_index: Optional[dict[str, int]] = None
|
||||
|
||||
@property
|
||||
def req_ids(self) -> list[str]:
|
||||
# None elements should only be present transiently
|
||||
@@ -407,6 +435,9 @@ class InputBatch:
|
||||
else:
|
||||
raise NotImplementedError(request)
|
||||
|
||||
# Speculative decoding: by default 1 token is generated.
|
||||
self.num_accepted_tokens_cpu[req_index] = 1
|
||||
|
||||
# Add request lora ID
|
||||
if request.lora_request:
|
||||
lora_id = request.lora_request.lora_int_id
|
||||
@@ -506,6 +537,8 @@ class InputBatch:
|
||||
self.presence_penalties_cpu[i2], self.presence_penalties_cpu[i1]
|
||||
self.repetition_penalties_cpu[i1], self.repetition_penalties_cpu[i2] =\
|
||||
self.repetition_penalties_cpu[i2], self.repetition_penalties_cpu[i1]
|
||||
self.num_accepted_tokens_cpu[i1], self.num_accepted_tokens_cpu[i2] =\
|
||||
self.num_accepted_tokens_cpu[i2], self.num_accepted_tokens_cpu[i1]
|
||||
|
||||
# NOTE: the following is unsafe
|
||||
# self.token_ids_cpu[i1, ...], self.token_ids_cpu[i2, ...], =\
|
||||
@@ -612,6 +645,8 @@ class InputBatch:
|
||||
empty_index] = self.presence_penalties_cpu[last_req_index]
|
||||
self.repetition_penalties_cpu[
|
||||
empty_index] = self.repetition_penalties_cpu[last_req_index]
|
||||
self.num_accepted_tokens_cpu[
|
||||
empty_index] = self.num_accepted_tokens_cpu[last_req_index]
|
||||
generator = self.generators.pop(last_req_index, None)
|
||||
if generator is not None:
|
||||
self.generators[empty_index] = generator
|
||||
@@ -726,20 +761,13 @@ class InputBatch:
|
||||
pooling_params = [
|
||||
self.pooling_params[req_id] for req_id in self.req_ids
|
||||
]
|
||||
if vllm_version_is("0.10.1.1") or vllm_version_is("0.10.1"):
|
||||
return PoolingMetadata(
|
||||
prompt_lens=torch.from_numpy(
|
||||
self.num_prompt_tokens[:self.num_reqs]).to(self.device),
|
||||
prompt_token_ids=self.sampling_metadata.prompt_token_ids,
|
||||
pooling_params=pooling_params,
|
||||
)
|
||||
else:
|
||||
return PoolingMetadata(
|
||||
prompt_lens=torch.from_numpy(
|
||||
self.num_prompt_tokens[:self.num_reqs]),
|
||||
prompt_token_ids=self.sampling_metadata.prompt_token_ids,
|
||||
pooling_params=pooling_params,
|
||||
)
|
||||
|
||||
return PoolingMetadata(
|
||||
prompt_lens=torch.from_numpy(
|
||||
self.num_prompt_tokens[:self.num_reqs]),
|
||||
prompt_token_ids=self.sampling_metadata.prompt_token_ids,
|
||||
pooling_params=pooling_params,
|
||||
)
|
||||
|
||||
def _make_prompt_token_ids_tensor(self) -> torch.Tensor:
|
||||
max_prompt_len = self.num_prompt_tokens[:self.num_reqs].max()
|
||||
|
||||
@@ -18,18 +18,18 @@
|
||||
#
|
||||
|
||||
import copy
|
||||
from typing import Optional
|
||||
from typing import Optional, Union
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import torch_npu
|
||||
import vllm.envs as envs_vllm
|
||||
from torch_npu.op_plugin.atb._atb_ops import _register_atb_extensions
|
||||
from torch_npu.profiler import dynamic_profile as dp
|
||||
from vllm.config import VllmConfig
|
||||
from vllm.distributed import (ensure_model_parallel_initialized,
|
||||
init_distributed_environment)
|
||||
from vllm.distributed.kv_transfer import (ensure_kv_transfer_initialized,
|
||||
has_kv_transfer_group)
|
||||
from vllm.distributed.kv_transfer import ensure_kv_transfer_initialized
|
||||
from vllm.distributed.parallel_state import get_pp_group, get_tp_group
|
||||
from vllm.logger import logger
|
||||
from vllm.lora.request import LoRARequest
|
||||
@@ -38,22 +38,31 @@ from vllm.tasks import SupportedTask
|
||||
from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, GiB_bytes
|
||||
from vllm.v1.core.sched.output import SchedulerOutput
|
||||
from vllm.v1.kv_cache_interface import KVCacheConfig, KVCacheSpec
|
||||
from vllm.v1.outputs import EMPTY_MODEL_RUNNER_OUTPUT, ModelRunnerOutput
|
||||
from vllm.v1.outputs import (EMPTY_MODEL_RUNNER_OUTPUT, AsyncModelRunnerOutput,
|
||||
DraftTokenIds, ModelRunnerOutput)
|
||||
from vllm.v1.worker.worker_base import WorkerBase
|
||||
|
||||
from vllm_ascend.ascend_config import init_ascend_config
|
||||
import vllm_ascend.envs as envs_ascend
|
||||
from vllm_ascend.ascend_config import get_ascend_config, init_ascend_config
|
||||
from vllm_ascend.device_allocator.camem import CaMemAllocator
|
||||
from vllm_ascend.distributed.parallel_state import init_ascend_model_parallel
|
||||
from vllm_ascend.platform import NPUPlatform
|
||||
from vllm_ascend.utils import (init_ascend_soc_version,
|
||||
register_ascend_customop, sleep_mode_enabled,
|
||||
try_register_lib, vllm_version_is)
|
||||
try_register_lib)
|
||||
from vllm_ascend.worker.model_runner_v1 import NPUModelRunner
|
||||
|
||||
if not (vllm_version_is("0.10.1.1") or vllm_version_is("0.10.1")):
|
||||
from vllm.v1.outputs import DraftTokenIds
|
||||
else:
|
||||
DraftTokenIds = None
|
||||
torch._dynamo.trace_rules.clear_lru_cache() # noqa: E402
|
||||
from torch._dynamo.variables import TorchInGraphFunctionVariable # noqa: E402
|
||||
|
||||
torch_non_c_binding_in_graph_functions_npu = dict.fromkeys(
|
||||
["torch.npu.current_stream"],
|
||||
TorchInGraphFunctionVariable,
|
||||
) # noqa: E402
|
||||
torch_non_c_binding_in_graph_functions_npu[
|
||||
"torch.npu.stream"] = TorchInGraphFunctionVariable # noqa: E402
|
||||
torch._dynamo.trace_rules.torch_name_rule_map.append(
|
||||
torch_non_c_binding_in_graph_functions_npu) # noqa: E402
|
||||
|
||||
|
||||
class NPUWorker(WorkerBase):
|
||||
@@ -75,10 +84,21 @@ class NPUWorker(WorkerBase):
|
||||
from vllm_ascend import ops
|
||||
ops.register_dummy_fusion_op()
|
||||
_register_atb_extensions()
|
||||
register_ascend_customop()
|
||||
register_ascend_customop(vllm_config)
|
||||
# init ascend config and soc version
|
||||
init_ascend_config(vllm_config)
|
||||
init_ascend_soc_version()
|
||||
if get_ascend_config().use_sfa:
|
||||
# Direct import instead of using try_register_lib to ensure proper error handling when
|
||||
# custom_ops is necessary but not available (e.g., in DeepSeek v3.2 deployments)
|
||||
# yapf: disable
|
||||
import custom_ops # type: ignore # noqa
|
||||
|
||||
# yapf: enable
|
||||
logger.info(
|
||||
"custom_ops module loaded successfully. Custom operators like "
|
||||
"torch.ops.custom.npu_sparse_flash_attention are now available."
|
||||
)
|
||||
|
||||
super().__init__(vllm_config=vllm_config,
|
||||
local_rank=local_rank,
|
||||
@@ -103,6 +123,15 @@ class NPUWorker(WorkerBase):
|
||||
init_cached_hf_modules()
|
||||
|
||||
self.profiler = self._init_profiler()
|
||||
if sleep_mode_enabled():
|
||||
# Buffers saved before sleep
|
||||
self._sleep_saved_buffers: dict[str, torch.Tensor] = {}
|
||||
|
||||
# FixMe: this is a patch to fix the issue cause by https://github.com/vllm-project/vllm/commit/de94289a98d7ec52a5ef02719e01a1db8b505170
|
||||
from vllm.model_executor.layers.linear import \
|
||||
WEIGHT_LOADER_V2_SUPPORTED
|
||||
if "UnquantizedLinearMethod" in WEIGHT_LOADER_V2_SUPPORTED:
|
||||
WEIGHT_LOADER_V2_SUPPORTED.remove("UnquantizedLinearMethod")
|
||||
|
||||
def sleep(self, level: int = 1) -> None:
|
||||
if not sleep_mode_enabled():
|
||||
@@ -110,6 +139,13 @@ class NPUWorker(WorkerBase):
|
||||
"Sleep mode is not enabled. Please compile vllm-ascend with COMPILE_CUSTOM_KERNELS=1."
|
||||
)
|
||||
free_bytes_before_sleep = NPUPlatform.mem_get_info()[0]
|
||||
# Save the buffers before level 2 sleep
|
||||
if level == 2:
|
||||
model = self.model_runner.model
|
||||
self._sleep_saved_buffers = {
|
||||
name: buffer.cpu().clone()
|
||||
for name, buffer in model.named_buffers()
|
||||
}
|
||||
allocator = CaMemAllocator.get_instance()
|
||||
allocator.sleep(offload_tags=("weights", ) if level == 1 else tuple())
|
||||
free_bytes_after_sleep, total = NPUPlatform.mem_get_info()
|
||||
@@ -129,6 +165,14 @@ class NPUWorker(WorkerBase):
|
||||
allocator = CaMemAllocator.get_instance()
|
||||
allocator.wake_up(tags=tags)
|
||||
|
||||
# Restore the buffers after level 2 sleep
|
||||
if len(self._sleep_saved_buffers):
|
||||
model = self.model_runner.model
|
||||
for name, buffer in model.named_buffers():
|
||||
if name in self._sleep_saved_buffers:
|
||||
buffer.data.copy_(self._sleep_saved_buffers[name].data)
|
||||
self._sleep_saved_buffers = {}
|
||||
|
||||
def initialize_cache(self, num_gpu_blocks: int,
|
||||
num_cpu_blocks: int) -> None:
|
||||
self.cache_config.num_gpu_blocks = num_gpu_blocks
|
||||
@@ -195,36 +239,42 @@ class NPUWorker(WorkerBase):
|
||||
def execute_model(
|
||||
self,
|
||||
scheduler_output: "SchedulerOutput",
|
||||
) -> Optional[ModelRunnerOutput]:
|
||||
) -> Optional[Union[ModelRunnerOutput, AsyncModelRunnerOutput]]:
|
||||
# enable msMonitor to monitor the performance of vllm-ascend
|
||||
if envs_ascend.MSMONITOR_USE_DAEMON:
|
||||
dp.step()
|
||||
|
||||
intermediate_tensors = None
|
||||
if not get_pp_group().is_first_rank:
|
||||
forward_pass = scheduler_output.total_num_scheduled_tokens > 0
|
||||
if forward_pass and not get_pp_group().is_first_rank:
|
||||
intermediate_tensors = IntermediateTensors(
|
||||
get_pp_group().recv_tensor_dict(
|
||||
all_gather_group=get_tp_group()))
|
||||
|
||||
output = self.model_runner.execute_model(scheduler_output,
|
||||
intermediate_tensors)
|
||||
if isinstance(output, (ModelRunnerOutput, AsyncModelRunnerOutput)):
|
||||
return output
|
||||
|
||||
assert isinstance(output, IntermediateTensors)
|
||||
parallel_config = self.vllm_config.parallel_config
|
||||
if parallel_config.distributed_executor_backend != "external_launcher" \
|
||||
and not get_pp_group().is_last_rank:
|
||||
assert isinstance(output, IntermediateTensors)
|
||||
get_pp_group().send_tensor_dict(output.tensors,
|
||||
all_gather_group=get_tp_group())
|
||||
if not has_kv_transfer_group():
|
||||
return None
|
||||
assert parallel_config.distributed_executor_backend != (
|
||||
"external_launcher") and not get_pp_group().is_last_rank
|
||||
|
||||
kv_connector_output = output.kv_connector_output
|
||||
finished_sending = kv_connector_output.finished_sending
|
||||
finished_recving = kv_connector_output.finished_recving
|
||||
get_pp_group().send_tensor_dict(output.tensors,
|
||||
all_gather_group=get_tp_group())
|
||||
|
||||
if not finished_sending and not finished_recving:
|
||||
return EMPTY_MODEL_RUNNER_OUTPUT
|
||||
kv_connector_output = output.kv_connector_output
|
||||
if not kv_connector_output:
|
||||
return None
|
||||
|
||||
new_output = copy.copy(EMPTY_MODEL_RUNNER_OUTPUT)
|
||||
new_output.kv_connector_output = kv_connector_output
|
||||
return new_output
|
||||
|
||||
assert isinstance(output, ModelRunnerOutput)
|
||||
# In case of PP with kv transfer, we need to pass through the
|
||||
# kv_connector_output
|
||||
if (not kv_connector_output.finished_sending
|
||||
and not kv_connector_output.finished_recving):
|
||||
return EMPTY_MODEL_RUNNER_OUTPUT
|
||||
output = copy.copy(EMPTY_MODEL_RUNNER_OUTPUT)
|
||||
output.kv_connector_output = kv_connector_output
|
||||
return output
|
||||
|
||||
def load_model(self) -> None:
|
||||
@@ -242,6 +292,7 @@ class NPUWorker(WorkerBase):
|
||||
|
||||
def compile_or_warm_up_model(self) -> None:
|
||||
# Note: need to adapt for graph mode.
|
||||
self.model_runner.eplb_warmup()
|
||||
warmup_sizes = (self.vllm_config.compilation_config.compile_sizes
|
||||
or []).copy()
|
||||
if not self.model_config.enforce_eager:
|
||||
@@ -254,10 +305,19 @@ class NPUWorker(WorkerBase):
|
||||
self.model_runner._dummy_run(size)
|
||||
if not self.model_config.enforce_eager:
|
||||
self.model_runner.capture_model()
|
||||
# Call ATB matmul to warm up; otherwise, the first operation (ReshapeAndCache)
|
||||
# may cause performance degradation at runtime.
|
||||
self._warm_up_atb()
|
||||
# Reset the seed to ensure that the random state is not affected by
|
||||
# the model initialization and profiling.
|
||||
NPUPlatform.seed_everything(self.model_config.seed)
|
||||
|
||||
def _warm_up_atb(self):
|
||||
x = torch.rand((2, 4), dtype=torch.float16).npu()
|
||||
weight = torch.rand((2, 4), dtype=torch.float16).npu()
|
||||
c = torch.rand((4, 4), dtype=torch.float32).npu()
|
||||
torch_npu._npu_matmul_add_fp32(x, weight, c)
|
||||
|
||||
def get_model(self) -> nn.Module:
|
||||
return self.model_runner.get_model()
|
||||
|
||||
@@ -313,6 +373,10 @@ class NPUWorker(WorkerBase):
|
||||
# Torch profiler. Enabled and configured through env vars:
|
||||
# VLLM_TORCH_PROFILER_DIR=/path/to/save/trace
|
||||
if envs_vllm.VLLM_TORCH_PROFILER_DIR:
|
||||
if envs_ascend.MSMONITOR_USE_DAEMON:
|
||||
raise RuntimeError(
|
||||
"MSMONITOR_USE_DAEMON and VLLM_TORCH_PROFILER_DIR cannot be both set at the same time."
|
||||
)
|
||||
torch_profiler_trace_dir = envs_vllm.VLLM_TORCH_PROFILER_DIR
|
||||
logger.info("Profiling enabled. Traces will be saved to: %s",
|
||||
torch_profiler_trace_dir)
|
||||
|
||||
Reference in New Issue
Block a user