v0.10.1rc1

This commit is contained in:
2025-09-09 09:40:35 +08:00
parent d6f6ef41fe
commit 9149384e03
432 changed files with 84698 additions and 1 deletions

View File

View File

@@ -0,0 +1,398 @@
# 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

View File

@@ -0,0 +1,439 @@
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

View File

@@ -0,0 +1,821 @@
#
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
# Copyright 2023 The vLLM team.
#
# 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.
# Adapted from vllm-project/vllm/vllm/worker/gpu_input_batch.py
#
from dataclasses import dataclass
from typing import Optional, cast
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.pooling_params import PoolingParams
from vllm.sampling_params import SamplingParams, SamplingType
from vllm.utils import swap_dict_values
from vllm.v1.outputs import LogprobsTensors
from vllm.v1.pool.metadata import PoolingMetadata
from vllm.v1.sample.logits_processor import (BatchUpdateBuilder,
LogitsProcessors,
MoveDirectionality)
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
@dataclass
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]
block_ids: tuple[list[int], ...]
num_computed_tokens: int
output_token_ids: list[int]
mrope_positions: Optional[torch.Tensor] = None
mrope_position_delta: Optional[int] = None
lora_request: Optional[LoRARequest] = None
def __post_init__(self):
self.num_prompt_tokens = len(self.prompt_token_ids)
@property
def num_tokens(self) -> int:
return self.num_prompt_tokens + len(self.output_token_ids)
# Temporary back-compatibility for plugins that define model runner
@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 get_token_id(self, idx: int) -> int:
if idx < self.num_prompt_tokens:
return self.prompt_token_ids[idx]
else:
return self.output_token_ids[idx - self.num_prompt_tokens]
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.is_pooling_model = is_pooling_model
self.is_spec_decode = is_spec_decode
self.max_num_reqs = max_num_reqs
self.max_model_len = max_model_len
self.max_num_batched_tokens = max_num_batched_tokens
self.device = device
self.pin_memory = pin_memory
self.vocab_size = vocab_size
self._req_ids: list[Optional[str]] = []
self.req_id_to_index: dict[str, int] = {}
# TODO(woosuk): This buffer could be too large if max_model_len is big.
# Find a way to reduce the CPU memory usage.
# This buffer is not directly transferred to the NPU, so it does not
# need to be pinned.
self.token_ids_cpu_tensor = torch.zeros(
(max_num_reqs, max_model_len),
device="cpu",
dtype=torch.int32,
pin_memory=False,
)
self.token_ids_cpu = self.token_ids_cpu_tensor.numpy()
self.num_tokens = np.zeros(max_num_reqs, dtype=np.int32)
self.num_tokens_no_spec = np.zeros(max_num_reqs, dtype=np.int32)
self.num_prompt_tokens = np.zeros(max_num_reqs, dtype=np.int32)
self.num_computed_tokens_cpu_tensor = torch.zeros(
(max_num_reqs, ),
device="cpu",
dtype=torch.int32,
pin_memory=pin_memory,
)
self.num_computed_tokens_cpu = \
self.num_computed_tokens_cpu_tensor.numpy()
# Block table.
self.block_table = MultiGroupBlockTable(
max_num_reqs=max_num_reqs,
max_model_len=max_model_len,
max_num_batched_tokens=max_num_batched_tokens,
pin_memory=pin_memory,
device=device,
block_sizes=block_sizes,
)
# Sampling-related.
self.temperature = torch.empty((max_num_reqs, ),
dtype=torch.float32,
device=device)
self.temperature_cpu_tensor = torch.empty((max_num_reqs, ),
dtype=torch.float32,
device="cpu",
pin_memory=pin_memory)
self.temperature_cpu = self.temperature_cpu_tensor.numpy()
self.greedy_reqs: set[str] = set()
self.random_reqs: set[str] = set()
self.top_p = torch.empty((max_num_reqs, ),
dtype=torch.float32,
device=device)
self.top_p_cpu_tensor = torch.empty((max_num_reqs, ),
dtype=torch.float32,
device="cpu",
pin_memory=pin_memory)
self.top_p_cpu = self.top_p_cpu_tensor.numpy()
self.top_p_reqs: set[str] = set()
self.top_k = torch.empty((max_num_reqs, ),
dtype=torch.int32,
device=device)
self.top_k_cpu_tensor = torch.empty((max_num_reqs, ),
dtype=torch.int32,
device="cpu",
pin_memory=pin_memory)
self.top_k_cpu = self.top_k_cpu_tensor.numpy()
self.top_k_reqs: set[str] = set()
# IDs of requests which do not support spec decoding
self.spec_decode_unsupported_reqs: set[str] = set()
# Frequency penalty related data structures
self.frequency_penalties = torch.empty((max_num_reqs, ),
dtype=torch.float,
device=device)
self.frequency_penalties_cpu_tensor = torch.empty(
(max_num_reqs, ),
dtype=torch.float,
device="cpu",
pin_memory=pin_memory)
self.frequency_penalties_cpu = \
self.frequency_penalties_cpu_tensor.numpy()
self.frequency_penalties_reqs: set[str] = set()
# Presence penalty related data structures
self.presence_penalties = torch.empty((max_num_reqs, ),
dtype=torch.float,
device=device)
self.presence_penalties_cpu_tensor = torch.empty((max_num_reqs, ),
dtype=torch.float,
device="cpu",
pin_memory=pin_memory)
self.presence_penalties_cpu = self.presence_penalties_cpu_tensor.numpy(
)
self.presence_penalties_reqs: set[str] = set()
# Repetition penalty related data structures
self.repetition_penalties = torch.empty((max_num_reqs, ),
dtype=torch.float,
device=device)
self.repetition_penalties_cpu_tensor = torch.empty(
(max_num_reqs, ),
dtype=torch.float,
device="cpu",
pin_memory=pin_memory)
self.repetition_penalties_cpu = \
self.repetition_penalties_cpu_tensor.numpy()
self.repetition_penalties_reqs: set[str] = set()
# lora related
self.request_lora_mapping = np.zeros((self.max_num_reqs, ),
dtype=np.int32)
self.lora_id_to_request_ids: dict[int, set[str]] = {}
self.lora_id_to_lora_request: dict[int, LoRARequest] = {}
# req_index -> generator
# NOTE(woosuk): The indices of the requests that do not have their own
# generator should not be included in the dictionary.
self.generators: dict[int, torch.Generator] = {}
self.num_logprobs: dict[str, int] = {}
# NOTE(rob): num_prompt_logprobs only includes reqs
# that are currently in the prefill phase.
self.num_prompt_logprobs: dict[str, int] = {}
# To accumulate prompt logprobs tensor chunks across prefill steps.
self.in_progress_prompt_logprobs_cpu: dict[str, LogprobsTensors] = {}
# Internal representation of per-step batch state changes, used for
# reordering persistent batch and generating logitsprocs batch state
# updates. Should reset each step.
self.batch_update_builder = BatchUpdateBuilder()
# TODO convert this to LogitsProcessor
self.has_allowed_token_ids: set[str] = set()
# NOTE(lufang): In the mask tensor, if the corresponding token allowed,
# the value is False. Since we use masked_fill_ to set -inf.
self.allowed_token_ids_mask: Optional[torch.Tensor] = None
self.allowed_token_ids_mask_cpu_tensor: Optional[torch.Tensor] = None
# req_index -> bad_words_token_ids
self.bad_words_token_ids: dict[int, list[list[int]]] = {}
self.logits_processing_needs_token_ids = np.zeros(max_num_reqs,
dtype=bool)
self.req_output_token_ids: list[Optional[list[int]]] = []
# Store provided logitsprocs. If none are provided, initialize empty
# data structure
self.logitsprocs = logitsprocs or LogitsProcessors()
# This is updated each time the batch constituents change.
self.sampling_metadata = self._make_sampling_metadata()
self.pooling_params: dict[str, PoolingParams] = {}
@property
def req_ids(self) -> list[str]:
# None elements should only be present transiently
# while performing state updates to the batch.
return cast(list[str], self._req_ids)
def _register_add_request(self, request: "CachedRequestState") -> int:
"""Track add-request operations for logits processors.
Not applicable to pooling models.
"""
# Detailed added request metadata is only required for non-pooling
# models, to support logitsprocs
assert request.sampling_params
# Fill the next empty index if there is one.
if (new_req_index := self.batch_update_builder.pop_removed()) is None:
# Append to end otherwise.
new_req_index = self.num_reqs
assert new_req_index < self.max_num_reqs
self.batch_update_builder.added.append(
(new_req_index, request.sampling_params, request.prompt_token_ids,
request.output_token_ids))
return new_req_index
def add_request(
self,
request: "CachedRequestState",
) -> int:
if not self.is_pooling_model:
# New request index bookkeeping for autoregressive models.
req_index = self._register_add_request(request)
else:
req_index = self.num_reqs
req_id = request.req_id
if req_index == len(self._req_ids):
self._req_ids.append(req_id)
self.req_output_token_ids.append(request.output_token_ids)
else:
self._req_ids[req_index] = req_id
self.req_output_token_ids[req_index] = request.output_token_ids
self.req_id_to_index[req_id] = req_index
# Copy the prompt token ids and output token ids.
num_prompt_tokens = len(request.prompt_token_ids)
self.num_prompt_tokens[req_index] = num_prompt_tokens
self.token_ids_cpu[
req_index, :num_prompt_tokens] = request.prompt_token_ids
start_idx = num_prompt_tokens
end_idx = start_idx + len(request.output_token_ids)
self.token_ids_cpu[req_index,
start_idx:end_idx] = request.output_token_ids
# Number of token ids in token_ids_cpu.
# NOTE(woosuk): This may include spec decode tokens.
self.num_tokens[req_index] = request.num_tokens
# Number of tokens without spec decode tokens.
self.num_tokens_no_spec[req_index] = request.num_tokens
self.num_computed_tokens_cpu[req_index] = request.num_computed_tokens
self.block_table.add_row(request.block_ids, req_index)
if sampling_params := request.sampling_params:
if (self.is_spec_decode
and is_spec_decode_unsupported(sampling_params)):
self.spec_decode_unsupported_reqs.add(req_id)
if sampling_params.sampling_type == SamplingType.GREEDY:
# Avoid later division by zero.
self.temperature_cpu[req_index] = -1.0
self.greedy_reqs.add(req_id)
else:
self.temperature_cpu[req_index] = sampling_params.temperature
self.random_reqs.add(req_id)
self.top_p_cpu[req_index] = sampling_params.top_p
if sampling_params.top_p < 1:
self.top_p_reqs.add(req_id)
top_k = sampling_params.top_k
if 0 < top_k < self.vocab_size:
self.top_k_reqs.add(req_id)
else:
top_k = self.vocab_size
self.top_k_cpu[req_index] = top_k
self.frequency_penalties_cpu[
req_index] = sampling_params.frequency_penalty
if sampling_params.frequency_penalty != 0.0:
self.frequency_penalties_reqs.add(req_id)
self.presence_penalties_cpu[
req_index] = sampling_params.presence_penalty
if sampling_params.presence_penalty != 0.0:
self.presence_penalties_reqs.add(req_id)
self.repetition_penalties_cpu[
req_index] = sampling_params.repetition_penalty
if sampling_params.repetition_penalty != 1.0:
self.repetition_penalties_reqs.add(req_id)
# NOTE(woosuk): self.generators should not include the requests that
# do not have their own generator.
if request.generator is not None:
self.generators[req_index] = request.generator
if sampling_params.logprobs is not None:
self.num_logprobs[req_id] = (self.vocab_size
if sampling_params.logprobs == -1
else sampling_params.logprobs)
if sampling_params.prompt_logprobs is not None:
self.num_prompt_logprobs[
req_id] = sampling_params.prompt_logprobs
if sampling_params.allowed_token_ids:
self.has_allowed_token_ids.add(req_id)
if self.allowed_token_ids_mask_cpu_tensor is None:
# Lazy allocation for this tensor, which can be large.
# False means we don't fill with -inf.
self.allowed_token_ids_mask = torch.zeros(
self.max_num_reqs,
self.vocab_size,
dtype=torch.bool,
device=self.device)
self.allowed_token_ids_mask_cpu_tensor = torch.zeros(
self.max_num_reqs,
self.vocab_size,
dtype=torch.bool,
device="cpu")
self.allowed_token_ids_mask_cpu_tensor[req_index] = True
# False means we don't fill with -inf.
self.allowed_token_ids_mask_cpu_tensor[req_index][
sampling_params.allowed_token_ids] = False
if sampling_params.bad_words_token_ids:
self.bad_words_token_ids[
req_index] = sampling_params.bad_words_token_ids
elif pooling_params := request.pooling_params:
self.pooling_params[req_id] = pooling_params
self.logits_processing_needs_token_ids[req_index] = (
pooling_params.requires_token_ids)
else:
raise NotImplementedError(request)
# Add request lora ID
if request.lora_request:
lora_id = request.lora_request.lora_int_id
if lora_id not in self.lora_id_to_request_ids:
self.lora_id_to_request_ids[lora_id] = set()
self.request_lora_mapping[req_index] = lora_id
self.lora_id_to_request_ids[lora_id].add(request.req_id)
self.lora_id_to_lora_request[lora_id] = request.lora_request
else:
# No LoRA
self.request_lora_mapping[req_index] = 0
return req_index
def remove_request(self, req_id: str) -> Optional[int]:
"""This method must always be followed by a call to condense().
Args:
req_id: request to remove
Returns:
Removed request index, or `None` if `req_id` not recognized
"""
req_index = self.req_id_to_index.pop(req_id, None)
if req_index is None:
return None
if not self.is_pooling_model:
# Autoregressive models require bookkeeping of removed requests to
# support logitsprocs.
self.batch_update_builder.removed_append(req_index)
self._req_ids[req_index] = None
self.req_output_token_ids[req_index] = None
self.greedy_reqs.discard(req_id)
self.random_reqs.discard(req_id)
self.top_p_reqs.discard(req_id)
self.top_k_reqs.discard(req_id)
self.spec_decode_unsupported_reqs.discard(req_id)
self.frequency_penalties_reqs.discard(req_id)
self.presence_penalties_reqs.discard(req_id)
self.repetition_penalties_reqs.discard(req_id)
self.generators.pop(req_index, None)
self.num_logprobs.pop(req_id, None)
self.num_prompt_logprobs.pop(req_id, None)
self.in_progress_prompt_logprobs_cpu.pop(req_id, None)
# LoRA
lora_id = self.request_lora_mapping[req_index]
if lora_id != 0:
self.lora_id_to_request_ids[lora_id].discard(req_id)
if len(self.lora_id_to_request_ids[lora_id]) == 0:
self.lora_id_to_request_ids.pop(lora_id)
self.lora_id_to_lora_request.pop(lora_id)
self.request_lora_mapping[req_index] = 0
self.has_allowed_token_ids.discard(req_id)
if self.allowed_token_ids_mask_cpu_tensor is not None:
# False means we don't fill with -inf.
self.allowed_token_ids_mask_cpu_tensor[req_index].fill_(False)
self.bad_words_token_ids.pop(req_index, None)
self.pooling_params.pop(req_id, None)
return req_index
def swap_states(self, i1: int, i2: int) -> None:
# For autoregressive models, track detailed request reordering info
# to support logitsprocs
self.batch_update_builder.moved.append(
(i1, i2, MoveDirectionality.SWAP))
old_id_i1 = self._req_ids[i1]
old_id_i2 = self._req_ids[i2]
self._req_ids[i1], self._req_ids[i2] =\
self._req_ids[i2], self._req_ids[i1] # noqa
self.req_output_token_ids[i1], self.req_output_token_ids[i2] =\
self.req_output_token_ids[i2], self.req_output_token_ids[i1]
assert old_id_i1 is not None and old_id_i2 is not None
self.req_id_to_index[old_id_i1], self.req_id_to_index[old_id_i2] =\
self.req_id_to_index[old_id_i2], self.req_id_to_index[old_id_i1]
self.num_tokens[i1], self.num_tokens[i2] =\
self.num_tokens[i2], self.num_tokens[i1]
self.num_tokens_no_spec[i1], self.num_tokens_no_spec[i2] =\
self.num_tokens_no_spec[i2], self.num_tokens_no_spec[i1]
self.num_prompt_tokens[i1], self.num_prompt_tokens[i2] =\
self.num_prompt_tokens[i2], self.num_prompt_tokens[i1]
self.num_computed_tokens_cpu[i1], self.num_computed_tokens_cpu[i2] =\
self.num_computed_tokens_cpu[i2], self.num_computed_tokens_cpu[i1]
self.temperature_cpu[i1], self.temperature_cpu[i2] =\
self.temperature_cpu[i2], self.temperature_cpu[i1]
self.top_p_cpu[i1], self.top_p_cpu[i2] =\
self.top_p_cpu[i2], self.top_p_cpu[i1]
self.top_k_cpu[i1], self.top_k_cpu[i2] =\
self.top_k_cpu[i2], self.top_k_cpu[i1]
self.frequency_penalties_cpu[i1], self.frequency_penalties_cpu[i2] =\
self.frequency_penalties_cpu[i2], self.frequency_penalties_cpu[i1]
self.presence_penalties_cpu[i1], self.presence_penalties_cpu[i2] =\
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]
# NOTE: the following is unsafe
# self.token_ids_cpu[i1, ...], self.token_ids_cpu[i2, ...], =\
# self.token_ids_cpu[i2, ...], self.token_ids_cpu[i1, ...]
# instead, we need to temporiarily copy the data for one of the indices
# TODO(lucas): optimize this by only copying valid indices
tmp = self.token_ids_cpu[i1, ...].copy()
self.token_ids_cpu[i1, ...] = self.token_ids_cpu[i2, ...]
self.token_ids_cpu[i2, ...] = tmp
swap_dict_values(self.generators, i1, i2)
swap_dict_values(self.bad_words_token_ids, i1, i2)
self.request_lora_mapping[i1], self.request_lora_mapping[i2] =\
self.request_lora_mapping[i2], self.request_lora_mapping[i1]
if self.allowed_token_ids_mask_cpu_tensor is not None:
self.allowed_token_ids_mask_cpu_tensor[i1], \
self.allowed_token_ids_mask_cpu_tensor[i2] =\
self.allowed_token_ids_mask_cpu_tensor[i2], \
self.allowed_token_ids_mask_cpu_tensor[i1]
self.block_table.swap_row(i1, i2)
def condense(self) -> None:
"""Slide non-empty requests down into lower, empty indices.
Any consecutive empty indices at the very end of the list are not
filled.
Args:
empty_req_indices: empty indices which may be filled.
Returns:
swaps: list of (from,to) swap tuples for moved requests
empty_req_indices: indices not filled by condensation
"""
num_reqs = self.num_reqs
if self.is_pooling_model:
# Will be contiguous in pooling case, just trim the lists.
del self._req_ids[num_reqs:]
del self.req_output_token_ids[num_reqs:]
return
if not (empty_req_indices := self.batch_update_builder.removed):
# All removed requests were replaced by added requests, or else no
# requests were removed at all. No condense() needed
return
if num_reqs == 0:
# The batched states are empty.
self._req_ids.clear()
self.req_output_token_ids.clear()
return
# NOTE(woosuk): This function assumes that the empty_req_indices
# is sorted in descending order.
last_req_index = num_reqs + len(empty_req_indices) - 1
while empty_req_indices:
# Find the largest non-empty index.
while last_req_index in empty_req_indices:
last_req_index -= 1
# Find the smallest empty index.
empty_index = self.batch_update_builder.peek_removed()
assert empty_index is not None
if empty_index >= last_req_index:
break
# Move active request down into empty request
# index.
self.batch_update_builder.pop_removed()
# Autoregressive models require detailed tracking of condense
# operations to support logitsprocs
self.batch_update_builder.moved.append(
(last_req_index, empty_index,
MoveDirectionality.UNIDIRECTIONAL))
req_id = self._req_ids[last_req_index]
output_token_ids = self.req_output_token_ids[last_req_index]
assert req_id is not None
self._req_ids[empty_index] = req_id
self._req_ids[last_req_index] = None
self.req_output_token_ids[empty_index] = output_token_ids
self.req_output_token_ids[last_req_index] = None
self.req_id_to_index[req_id] = empty_index
num_tokens = self.num_tokens[last_req_index]
self.token_ids_cpu[empty_index, :num_tokens] = self.token_ids_cpu[
last_req_index, :num_tokens]
self.num_tokens[empty_index] = num_tokens
self.num_tokens_no_spec[empty_index] = self.num_tokens_no_spec[
last_req_index]
self.num_prompt_tokens[empty_index] = self.num_prompt_tokens[
last_req_index]
self.num_computed_tokens_cpu[
empty_index] = self.num_computed_tokens_cpu[last_req_index]
self.block_table.move_row(last_req_index, empty_index)
self.temperature_cpu[empty_index] = self.temperature_cpu[
last_req_index]
self.top_p_cpu[empty_index] = self.top_p_cpu[last_req_index]
self.top_k_cpu[empty_index] = self.top_k_cpu[last_req_index]
self.frequency_penalties_cpu[
empty_index] = self.frequency_penalties_cpu[last_req_index]
self.presence_penalties_cpu[
empty_index] = self.presence_penalties_cpu[last_req_index]
self.repetition_penalties_cpu[
empty_index] = self.repetition_penalties_cpu[last_req_index]
generator = self.generators.pop(last_req_index, None)
if generator is not None:
self.generators[empty_index] = generator
self.request_lora_mapping[empty_index] = self.request_lora_mapping[
last_req_index]
# TODO convert these to LogitsProcessors
if self.allowed_token_ids_mask_cpu_tensor is not None:
self.allowed_token_ids_mask_cpu_tensor[
empty_index] = self.allowed_token_ids_mask_cpu_tensor[
last_req_index]
bad_words_token_ids = self.bad_words_token_ids.pop(
last_req_index, None)
if bad_words_token_ids is not None:
self.bad_words_token_ids[empty_index] = bad_words_token_ids
# Decrement last_req_index since it is now empty.
last_req_index -= 1
# Trim lists to the batch size.
del self._req_ids[num_reqs:]
del self.req_output_token_ids[num_reqs:]
def refresh_metadata(self):
"""Apply any batch updates to sampling metadata."""
if self.is_pooling_model:
# Batch changes every step for pooling models.
self.sampling_metadata = self._make_sampling_metadata()
return
# For non-pooling models - generate and apply logitsprocs update;
# reset batch update tracking.
# Update sampling metadata if batch state is changed.
batch_update = self.batch_update_builder.get_and_reset(self.num_reqs)
for logit_proc in self.logitsprocs.all:
logit_proc.update_state(batch_update)
if batch_update:
self.sampling_metadata = self._make_sampling_metadata()
def _make_sampling_metadata(self) -> SamplingMetadata:
num_reqs = self.num_reqs
if not self.all_greedy:
temperature = copy_slice(self.temperature_cpu_tensor,
self.temperature, num_reqs)
else:
temperature = None
if not self.no_top_p:
copy_slice(self.top_p_cpu_tensor, self.top_p, num_reqs)
if not self.no_top_k:
copy_slice(self.top_k_cpu_tensor, self.top_k, num_reqs)
if not self.no_penalties:
# Since syncing these tensors is expensive only copy them
# if necessary i.e. if there are requests which require
# penalties to be applied during sampling.
copy_slice(self.frequency_penalties_cpu_tensor,
self.frequency_penalties, num_reqs)
copy_slice(self.presence_penalties_cpu_tensor,
self.presence_penalties, num_reqs)
copy_slice(self.repetition_penalties_cpu_tensor,
self.repetition_penalties, num_reqs)
needs_prompt_token_ids = (
not self.no_penalties
or self.logits_processing_needs_token_ids[:num_reqs].any())
if needs_prompt_token_ids:
# The prompt tokens are used only for applying penalties or
# step pooling during the sampling/pooling process.
# Hence copy these tensors only when there are requests which
# need penalties/step_pooler to be applied.
prompt_token_ids = self._make_prompt_token_ids_tensor()
else:
prompt_token_ids = None
allowed_token_ids_mask: Optional[torch.Tensor] = None
if not self.no_allowed_token_ids:
assert self.allowed_token_ids_mask is not None
copy_slice(self.allowed_token_ids_mask_cpu_tensor,
self.allowed_token_ids_mask, num_reqs)
allowed_token_ids_mask = self.allowed_token_ids_mask[:num_reqs]
return SamplingMetadata(
temperature=temperature,
all_greedy=self.all_greedy,
all_random=self.all_random,
top_p=None if self.no_top_p else self.top_p[:num_reqs],
top_k=None if self.no_top_k else self.top_k[:num_reqs],
generators=self.generators,
max_num_logprobs=self.max_num_logprobs,
prompt_token_ids=prompt_token_ids,
frequency_penalties=self.frequency_penalties[:num_reqs],
presence_penalties=self.presence_penalties[:num_reqs],
repetition_penalties=self.repetition_penalties[:num_reqs],
output_token_ids=cast(list[list[int]], self.req_output_token_ids),
no_penalties=self.no_penalties,
allowed_token_ids_mask=allowed_token_ids_mask,
bad_words_token_ids=self.bad_words_token_ids,
logitsprocs=self.logitsprocs,
)
@property
def pooling_metadata(self) -> PoolingMetadata:
if len(self.pooling_params) == 0:
pooling_params = []
else:
# Note, for now this assumes that all request in the batch
# are either sampling or pooling requests
assert len(self.req_ids) == len(self.pooling_params)
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,
)
def _make_prompt_token_ids_tensor(self) -> torch.Tensor:
max_prompt_len = self.num_prompt_tokens[:self.num_reqs].max()
prompt_token_ids_cpu_tensor = torch.empty(
(self.num_reqs, max_prompt_len),
device="cpu",
dtype=torch.int64,
pin_memory=self.pin_memory,
)
prompt_token_ids = prompt_token_ids_cpu_tensor.numpy()
prompt_token_ids[:] = self.token_ids_cpu[:self.
num_reqs, :max_prompt_len]
# Use the value of vocab_size as a pad since we don't have a
# token_id of this value.
for i in range(self.num_reqs):
prompt_token_ids[i, self.num_prompt_tokens[i]:] = self.vocab_size
return prompt_token_ids_cpu_tensor.to(device=self.device,
non_blocking=True)
def make_lora_inputs(
self, num_scheduled_tokens: np.ndarray
) -> tuple[tuple[int, ...], tuple[int, ...], set[LoRARequest]]:
"""
Given the num_scheduled_tokens for each request in the batch, return
datastructures used to activate the current LoRAs.
Returns:
1. prompt_lora_mapping: A tuple of size self.num_reqs where,
prompt_lora_mapping[i] is the LoRA id to use for the ith prompt.
2. token_lora_mapping: A tuple of size np.sum(num_scheduled_tokens)
where, token_lora_mapping[i] is the LoRA id to use for ith token.
3. lora_requests: Set of relevant LoRA requests.
"""
req_lora_mapping = self.request_lora_mapping[:self.num_reqs]
prompt_lora_mapping = tuple(req_lora_mapping)
token_lora_mapping = tuple(
req_lora_mapping.repeat(num_scheduled_tokens))
active_lora_requests: set[LoRARequest] = set(
self.lora_id_to_lora_request.values())
return prompt_lora_mapping, token_lora_mapping, active_lora_requests
@property
def num_reqs(self) -> int:
return len(self.req_id_to_index)
@property
def all_greedy(self) -> bool:
return len(self.random_reqs) == 0
@property
def all_random(self) -> bool:
return len(self.greedy_reqs) == 0
@property
def no_top_p(self) -> bool:
return len(self.top_p_reqs) == 0
@property
def no_top_k(self) -> bool:
return len(self.top_k_reqs) == 0
@property
def no_penalties(self) -> bool:
return (len(self.presence_penalties_reqs) == 0
and len(self.frequency_penalties_reqs) == 0
and len(self.repetition_penalties_reqs) == 0)
@property
def max_num_logprobs(self) -> Optional[int]:
return max(self.num_logprobs.values()) if self.num_logprobs else None
@property
def no_prompt_logprob(self) -> bool:
return not self.num_prompt_logprobs
@property
def no_allowed_token_ids(self) -> bool:
return len(self.has_allowed_token_ids) == 0

View File

@@ -0,0 +1,354 @@
#
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
# Copyright 2023 The vLLM team.
#
# 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.
# Adapted from vllm-project/vllm/vllm/worker/gpu_worker.py
#
import copy
from typing import Optional
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 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.parallel_state import get_pp_group, get_tp_group
from vllm.logger import logger
from vllm.lora.request import LoRARequest
from vllm.sequence import IntermediateTensors
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.worker.worker_base import WorkerBase
from vllm_ascend.ascend_config import 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)
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
class NPUWorker(WorkerBase):
def __init__(
self,
vllm_config: VllmConfig,
local_rank: int,
rank: int,
distributed_init_method: str,
is_driver_worker: bool = False,
# Additional parameters for compatibility with vllm
**kwargs):
"""Initialize the worker for Ascend."""
# register patch for vllm
from vllm_ascend.utils import adapt_patch
adapt_patch()
# Register ops when worker init.
from vllm_ascend import ops
ops.register_dummy_fusion_op()
_register_atb_extensions()
register_ascend_customop()
# init ascend config and soc version
init_ascend_config(vllm_config)
init_ascend_soc_version()
super().__init__(vllm_config=vllm_config,
local_rank=local_rank,
rank=rank,
distributed_init_method=distributed_init_method,
is_driver_worker=is_driver_worker)
# Try to import mindie_turbo to accelerate vLLM inference.
try_register_lib(
"mindie_turbo",
"MindIE Turbo is installed. vLLM inference will be accelerated with MindIE Turbo."
)
if self.cache_config.cache_dtype == "auto":
self.cache_dtype = self.model_config.dtype
else:
self.cache_dtype = STR_DTYPE_TO_TORCH_DTYPE[
self.cache_config.cache_dtype]
if self.model_config.trust_remote_code:
# note: lazy import to avoid importing torch before initializing
from vllm.utils import init_cached_hf_modules
init_cached_hf_modules()
self.profiler = self._init_profiler()
def sleep(self, level: int = 1) -> None:
if not sleep_mode_enabled():
raise ValueError(
"Sleep mode is not enabled. Please compile vllm-ascend with COMPILE_CUSTOM_KERNELS=1."
)
free_bytes_before_sleep = NPUPlatform.mem_get_info()[0]
allocator = CaMemAllocator.get_instance()
allocator.sleep(offload_tags=("weights", ) if level == 1 else tuple())
free_bytes_after_sleep, total = NPUPlatform.mem_get_info()
freed_bytes = free_bytes_after_sleep - free_bytes_before_sleep
used_bytes = total - free_bytes_after_sleep
assert freed_bytes >= 0, "Memory usage increased after sleeping."
logger.info(
"Sleep mode freed %.2f GiB memory, "
"%.2f GiB memory is still in use.", freed_bytes / GiB_bytes,
used_bytes / GiB_bytes)
def wake_up(self, tags: Optional[list[str]] = None) -> None:
if not sleep_mode_enabled():
raise ValueError(
"Sleep mode is not enabled. Please compile vllm-ascend with COMPILE_CUSTOM_KERNELS=1."
)
allocator = CaMemAllocator.get_instance()
allocator.wake_up(tags=tags)
def initialize_cache(self, num_gpu_blocks: int,
num_cpu_blocks: int) -> None:
self.cache_config.num_gpu_blocks = num_gpu_blocks
self.cache_config.num_cpu_blocks = num_cpu_blocks
def _init_device(self):
device = torch.device(f"npu:{self.local_rank}")
NPUPlatform.set_device(device)
NPUPlatform.empty_cache()
self.init_npu_memory = NPUPlatform.mem_get_info()[0]
# Initialize the distributed environment.
self._init_worker_distributed_environment()
# Set random seed.
NPUPlatform.seed_everything(self.model_config.seed)
return device
def init_device(self):
device = self._init_device()
# Init ModelRunner here, so that we have access to self.device.
self.model_runner = NPUModelRunner(self.vllm_config, device)
def determine_available_memory(self) -> int:
# Profile the memory usage of the model and get the maximum number of
# cache blocks that can be allocated with the remaining free memory.
NPUPlatform.clear_npu_memory()
# Execute a forward pass with dummy inputs to profile the memory usage
# of the model.
_, total_npu_memory = NPUPlatform.mem_get_info()
self.model_runner.profile_run()
# Calculate the number of blocks that can be allocated with the
# profiled peak memory.
free_npu_memory, _ = NPUPlatform.mem_get_info()
# NOTE(woosuk): Here we assume that the other processes using the same
# GPU did not change their memory usage during the profiling.
assert self.init_npu_memory > free_npu_memory, (
"Error in memory profiling. "
f"Initial free memory {self.init_npu_memory}, current free memory"
f" {free_npu_memory}. This happens when the NPU memory was "
"not properly cleaned up before initializing the vLLM instance.")
# Get the peak memory allocation recorded by torch
peak_memory = torch_npu.npu.memory_stats()["allocated_bytes.all.peak"]
# TODO: don`t need impl this func after empty_cache in
# Worker.determine_num_available_blocks() unified`
NPUPlatform.empty_cache()
torch_allocated_bytes = torch_npu.npu.memory_stats(
)["allocated_bytes.all.current"]
total_allocated_bytes = torch_npu.npu.mem_get_info(
)[1] - torch_npu.npu.mem_get_info()[0]
non_torch_allocations = total_allocated_bytes - torch_allocated_bytes
if non_torch_allocations > 0:
peak_memory += non_torch_allocations
available_kv_cache_memory = int(
total_npu_memory * self.cache_config.gpu_memory_utilization -
peak_memory)
available_kv_cache_memory = int(max(available_kv_cache_memory, 0))
logger.info(
f"Available memory: {available_kv_cache_memory}, total memory: {total_npu_memory}"
)
return available_kv_cache_memory
def execute_model(
self,
scheduler_output: "SchedulerOutput",
) -> Optional[ModelRunnerOutput]:
intermediate_tensors = None
if 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)
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
kv_connector_output = output.kv_connector_output
finished_sending = kv_connector_output.finished_sending
finished_recving = kv_connector_output.finished_recving
if not finished_sending and not finished_recving:
return EMPTY_MODEL_RUNNER_OUTPUT
new_output = copy.copy(EMPTY_MODEL_RUNNER_OUTPUT)
new_output.kv_connector_output = kv_connector_output
return new_output
assert isinstance(output, ModelRunnerOutput)
return output
def load_model(self) -> None:
if self.vllm_config.model_config.enable_sleep_mode:
allocator = CaMemAllocator.get_instance()
assert allocator.get_current_usage() == 0, (
"Sleep mode can only be "
"used for one instance per process.")
context = allocator.use_memory_pool(tag="weights")
else:
from contextlib import nullcontext
context = nullcontext() # type: ignore
with context:
self.model_runner.load_model()
def compile_or_warm_up_model(self) -> None:
# Note: need to adapt for graph mode.
warmup_sizes = (self.vllm_config.compilation_config.compile_sizes
or []).copy()
if not self.model_config.enforce_eager:
warmup_sizes = [
x for x in warmup_sizes if x not in
self.vllm_config.compilation_config.cudagraph_capture_sizes
]
for size in sorted(warmup_sizes, reverse=True):
logger.info("Compile and warming up model for size %d", size)
self.model_runner._dummy_run(size)
if not self.model_config.enforce_eager:
self.model_runner.capture_model()
# 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 get_model(self) -> nn.Module:
return self.model_runner.get_model()
def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
return self.model_runner.get_kv_cache_spec()
def initialize_from_config(self, kv_cache_config: KVCacheConfig) -> None:
"""Allocate NPU KV cache with the specified kv_cache_config."""
if self.vllm_config.model_config.enable_sleep_mode:
allocator = CaMemAllocator.get_instance()
context = allocator.use_memory_pool(tag="kv_cache")
else:
from contextlib import nullcontext
context = nullcontext() # type: ignore
with context:
self.model_runner.initialize_kv_cache(kv_cache_config)
def profile(self, is_start: bool = True):
if self.profiler is None:
raise RuntimeError("Profiler is not enabled.")
if is_start:
self.profiler.start()
else:
self.profiler.stop()
def add_lora(self, lora_request: LoRARequest) -> bool:
return self.model_runner.add_lora(lora_request)
def remove_lora(self, lora_id: int) -> bool:
return self.model_runner.remove_lora(lora_id)
def list_loras(self) -> set[int]:
return self.model_runner.list_loras()
def pin_lora(self, lora_id: int) -> bool:
return self.model_runner.pin_lora(lora_id)
def execute_dummy_batch(self) -> None:
self.model_runner._dummy_run(1)
def _init_worker_distributed_environment(self) -> None:
"""Initialize the distributed environment."""
init_distributed_environment(self.parallel_config.world_size,
self.rank, self.distributed_init_method,
self.local_rank, "hccl")
ensure_model_parallel_initialized(
self.parallel_config.tensor_parallel_size,
self.parallel_config.pipeline_parallel_size)
init_ascend_model_parallel(self.parallel_config)
ensure_kv_transfer_initialized(self.vllm_config)
def _init_profiler(self):
# Torch profiler. Enabled and configured through env vars:
# VLLM_TORCH_PROFILER_DIR=/path/to/save/trace
if envs_vllm.VLLM_TORCH_PROFILER_DIR:
torch_profiler_trace_dir = envs_vllm.VLLM_TORCH_PROFILER_DIR
logger.info("Profiling enabled. Traces will be saved to: %s",
torch_profiler_trace_dir)
experimental_config = torch_npu.profiler._ExperimentalConfig(
export_type=torch_npu.profiler.ExportType.Text,
profiler_level=torch_npu.profiler.ProfilerLevel.Level1,
msprof_tx=False,
aic_metrics=torch_npu.profiler.AiCMetrics.AiCoreNone,
l2_cache=False,
op_attr=False,
data_simplification=False,
record_op_args=False,
gc_detect_threshold=None,
)
return torch_npu.profiler.profile(
activities=[
torch_npu.profiler.ProfilerActivity.CPU,
torch_npu.profiler.ProfilerActivity.NPU,
],
with_stack=envs_vllm.VLLM_TORCH_PROFILER_WITH_STACK,
profile_memory=envs_vllm.\
VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY,
with_modules=False,
experimental_config=experimental_config,
on_trace_ready=torch_npu.profiler.tensorboard_trace_handler(
torch_profiler_trace_dir))
else:
return None
def get_supported_pooling_tasks(self):
return self.model_runner.get_supported_pooling_tasks()
def get_supported_tasks(self) -> "tuple[SupportedTask, ...]":
return self.model_runner.get_supported_tasks()
def take_draft_token_ids(self) -> Optional[DraftTokenIds]:
return self.model_runner.take_draft_token_ids()