EngineX-MetaX/enginex-c_series-vllm
9
1
Fork 1
Code Issues Pull Requests Actions Projects Releases Wiki Activity

[gpt-oss] Add gpt-oss bf16 support

Browse Source
This commit is contained in:
wangjing
2025-08-13 21:25:57 +08:00
parent 5d2e7edf78
commit 17ea2ec6aa
1232 changed files with 777 additions and 36 deletions
Download Patch File Download Diff File Expand all files Collapse all files

0
vllm/v1/spec_decode/__init__.py Normal file
View File

434
vllm/v1/spec_decode/eagle.py Normal file
View File

@@ -0,0 +1,434 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
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.forward_context import set_forward_context
from vllm.logger import init_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.attention.backends.flash_attn import (CommonAttentionMetadata,
FlashAttentionMetadata)
from vllm.v1.kv_cache_interface import KVCacheConfig
from vllm.v1.sample.metadata import SamplingMetadata
from vllm.v1.spec_decode.utils import prepare_eagle_input_kernel
logger = init_logger(__name__)
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.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)
# 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)
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:
num_tokens = target_token_ids.shape[0]
batch_size = next_token_ids.shape[0]
last_token_indices = cu_num_tokens[1:] - 1
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
# FA requires seq_len to have dtype int32.
seq_lens = (target_positions[last_token_indices] + 1).int()
if self.method in ["eagle", "eagle3"]:
# FIXME(woosuk): The below two ops cause synchronization. Optimize.
max_seq_len = seq_lens.max().item()
max_num_tokens = (cu_num_tokens[1:] -
cu_num_tokens[:-1]).max().item()
attn_metadata = FlashAttentionMetadata(
num_actual_tokens=num_tokens,
max_query_len=max_num_tokens,
query_start_loc=cu_num_tokens,
max_seq_len=max_seq_len,
seq_lens=seq_lens,
block_table=block_table,
slot_mapping=target_slot_mapping,
# TODO(woosuk): Support cascade attention.
use_cascade=False,
common_prefix_len=0,
cu_prefix_query_lens=None,
prefix_kv_lens=None,
suffix_kv_lens=None,
)
elif self.method == "deepseek_mtp":
query_lens = cu_num_tokens[1:] - cu_num_tokens[:-1]
max_query_len = query_lens.max().item()
common_attn_metadata = CommonAttentionMetadata(
query_start_loc=cu_num_tokens,
seq_lens=seq_lens,
num_reqs=batch_size,
num_actual_tokens=num_tokens,
max_query_len=max_query_len,
)
assert self.runner is not None
# FIXME: need to consider multiple kv_cache_groups
attn_metadata = self.runner.attn_metadata_builders[0].build(
common_prefix_len=0,
common_attn_metadata=common_attn_metadata,
)
else:
raise ValueError(f"Unsupported method: {self.method}")
# At this moment, we assume all eagle layers belong to the same KV
# cache group, thus using the same attention metadata.
per_layer_attn_metadata = {}
for layer_name in self.attn_layer_names:
per_layer_attn_metadata[layer_name] = attn_metadata
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
self.hidden_states[:num_tokens] = target_hidden_states
with set_forward_context(per_layer_attn_metadata,
self.vllm_config,
num_tokens=num_input_tokens):
ret_hidden_states = self.model(
self.input_ids[:num_input_tokens],
self.positions[:num_input_tokens],
self.hidden_states[:num_input_tokens],
)
if self.method == "deepseek_mtp":
last_hidden_states = ret_hidden_states
else:
last_hidden_states, hidden_states = ret_hidden_states
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)
# TODO: Currently, MTP module released by deepseek only has
# one layer. Adapt this code to support multiple layers once
# there's a multi-layer MTP module.
# Generate the remaining draft tokens.
draft_token_ids_list = [draft_token_ids]
positions = target_positions[last_token_indices]
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]
for _ 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_list[-1].int()
positions += 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 >= 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 = torch.where(exceeds_max_model_len, 0,
positions)
# Increment the sequence lengths.
attn_metadata.max_seq_len += 1
attn_metadata.seq_lens += 1
# 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
# sequence length to 1 to minimize their overheads in attention.
attn_metadata.seq_lens.masked_fill_(exceeds_max_model_len, 1)
# Compute the slot mapping.
block_numbers = clamped_positions // self.block_size
block_ids = block_table.gather(dim=1,
index=block_numbers.view(-1, 1))
block_ids = block_ids.view(-1)
attn_metadata.slot_mapping = (block_ids * self.block_size +
clamped_positions % 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.
attn_metadata.slot_mapping.masked_fill_(exceeds_max_model_len,
PADDING_SLOT_ID)
# 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
# Run the model.
with set_forward_context(per_layer_attn_metadata,
self.vllm_config,
num_tokens=input_batch_size):
last_hidden_states, hidden_states = self.model(
self.input_ids[:input_batch_size],
self.positions[:input_batch_size],
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_list.append(draft_token_ids)
# [batch_size, num_speculative_tokens]
draft_token_ids = torch.stack(draft_token_ids_list, dim=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,
)
batch_size = num_rejected_tokens.shape[0]
BLOCK_SIZE = 1024
prepare_eagle_input_kernel[(batch_size, )](
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)
# share embed_tokens with the target model if needed
if get_pp_group().world_size == 1 \
and self.model.model.embed_tokens.weight.shape \
== target_model.model.embed_tokens.weight.shape:
logger.info(
"Assuming the EAGLE head shares the same vocab embedding" \
" with the target model."
)
del self.model.model.embed_tokens
self.model.model.embed_tokens = target_model.model.embed_tokens
else:
logger.info(
"The EAGLE head's vocab embedding will be loaded separately" \
" from 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_forward_context(None, self.vllm_config,
num_tokens=num_tokens):
self.model(
self.input_ids[:num_tokens],
self.positions[:num_tokens],
self.hidden_states[:num_tokens],
)
def validate_same_kv_cache_group(self,
kv_cache_config: KVCacheConfig) -> None:
"""
Validate that all eagle layers belong to the same KVCacheGroup.
Need this assumption to ensure all eagle layers can use the
same AttentionMetadata.
May extend to multiple AttentionMetadata in the future.
"""
kv_cache_groups: dict[str, int] = {}
for id, kv_cache_group in enumerate(kv_cache_config.kv_cache_groups):
for layer_name in kv_cache_group.layer_names:
kv_cache_groups[layer_name] = id
assert len(
set([
kv_cache_groups[layer_name]
for layer_name in self.attn_layer_names
])
) == 1, "All eagle layers should belong to the same kv cache group"
# NOTE(woosuk): Currently, the below code is not used and we always use argmax
# to sample the draft tokens. We will use this after we find a way to manage
# the draft prob tensor.
# Refer to https://github.com/vllm-project/vllm/pull/16899 for the details.
# FIXME(woosuk): The logic here is duplicated with the main sampling code.
# We should refactor this to reuse the same sampling implementation.
def compute_probs_and_sample_next_token(
logits: torch.Tensor,
sampling_metadata: SamplingMetadata,
) -> tuple[torch.Tensor, torch.Tensor]:
if sampling_metadata.all_greedy:
# For greedy requests, draft_probs is not used in rejection sampling.
# Therefore, we can just return the logits.
probs = logits
next_token_ids = logits.argmax(dim=-1)
return next_token_ids, probs
is_greedy = sampling_metadata.temperature == -1
temperature = torch.where(is_greedy, 1.0, sampling_metadata.temperature)
logits.div_(temperature.view(-1, 1))
probs = logits.softmax(dim=-1, dtype=torch.float32)
# NOTE(woosuk): Currently, we ignore most of the sampling parameters in
# generating the draft tokens. We only use the temperature. While this
# could degrade the acceptance rate, it does not affect the distribution
# of the generated tokens after rejection sampling.
# TODO(woosuk): Consider seeds.
q = torch.empty_like(probs)
q.exponential_()
# NOTE(woosuk): We shouldn't use `probs.div_(q)` because the draft_probs
# will be used later for rejection sampling.
next_token_ids = probs.div(q).argmax(dim=-1).view(-1)
if not sampling_metadata.all_random:
greedy_token_ids = probs.argmax(dim=-1)
next_token_ids = torch.where(
is_greedy,
greedy_token_ids,
next_token_ids,
)
return next_token_ids, probs

62
vllm/v1/spec_decode/medusa.py Normal file
View File

@@ -0,0 +1,62 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import torch
import torch.nn as nn
from vllm.config import VllmConfig
from vllm.forward_context import set_forward_context
from vllm.logger import init_logger
from vllm.model_executor.model_loader import get_model
from vllm.v1.sample.metadata import SamplingMetadata
# Initialize logger
logger = init_logger(__name__)
class MedusaProposer:
"""
Medusa proposer class for generating token sequences
"""
def __init__(
self,
vllm_config: VllmConfig,
device: torch.device,
):
# Save config parameters
self.vllm_config = vllm_config
self.device = device
self.max_num_tokens = (
vllm_config.scheduler_config.max_num_batched_tokens)
self.hidden_size = vllm_config.speculative_config.\
draft_model_config.get_hidden_size(
)
self.dtype = vllm_config.model_config.dtype
def propose(
self,
target_hidden_states: torch.Tensor,
sampling_metadata: SamplingMetadata,
) -> torch.Tensor:
# Generate blocks and compute logits
blocks = self.model(target_hidden_states)
logits = self.model.compute_logits(blocks, None)
# Get draft tokens and transpose the result
draft_tokens = [logit.argmax(dim=-1).tolist() for logit in logits]
return [list(row) for row in zip(*draft_tokens)]
def load_model(self, target_model: nn.Module) -> None:
self.model = get_model(vllm_config=self.vllm_config,
model_config=self.vllm_config.
speculative_config.draft_model_config)
@torch.inference_mode()
def dummy_run(self, num_tokens: int) -> None:
hidden_states = torch.zeros((self.max_num_tokens, self.hidden_size),
dtype=self.dtype,
device=self.device)
with set_forward_context(None, self.vllm_config,
num_tokens=num_tokens):
self.model(hidden_states)

62
vllm/v1/spec_decode/metadata.py Normal file
View File

@@ -0,0 +1,62 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from dataclasses import dataclass
import numpy as np
import torch
@dataclass
class SpecDecodeMetadata:
# [num_tokens]
draft_token_ids: torch.Tensor
# [batch_size]
num_draft_tokens: list[int]
# [batch_size]
cu_num_draft_tokens: torch.Tensor
# [num_tokens]
target_logits_indices: torch.Tensor
# [batch_size]
bonus_logits_indices: torch.Tensor
# [num_tokens + batch_size]
logits_indices: torch.Tensor
def __post_init__(self):
self.max_spec_len = max(self.num_draft_tokens)
@classmethod
def make_dummy(
cls,
draft_token_ids: list[list[int]],
device: torch.device,
) -> "SpecDecodeMetadata":
batch_size = len(draft_token_ids)
num_draft_tokens = [len(ids) for ids in draft_token_ids]
flattened_draft_token_ids = sum(draft_token_ids, [])
num_tokens = len(flattened_draft_token_ids)
draft_token_ids_tensor = torch.tensor(flattened_draft_token_ids,
dtype=torch.int32,
device=device)
cu_num_draft_tokens = np.cumsum(num_draft_tokens, dtype=np.int32)
cu_num_draft_tokens_tensor = torch.from_numpy(cu_num_draft_tokens).to(
device)
target_logits_indices = torch.zeros(num_tokens,
dtype=torch.int32,
device=device)
bonus_logits_indices = torch.zeros(batch_size,
dtype=torch.int32,
device=device)
logits_indices = torch.zeros(num_tokens + batch_size,
dtype=torch.int32,
device=device)
return cls(
draft_token_ids=draft_token_ids_tensor,
num_draft_tokens=num_draft_tokens,
cu_num_draft_tokens=cu_num_draft_tokens_tensor,
target_logits_indices=target_logits_indices,
bonus_logits_indices=bonus_logits_indices,
logits_indices=logits_indices,
)

178
vllm/v1/spec_decode/metrics.py Normal file
View File

@@ -0,0 +1,178 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from dataclasses import dataclass, field
from typing import Optional
import numpy as np
import prometheus_client
from vllm.config import SpeculativeConfig
from vllm.logger import init_logger
logger = init_logger(__name__)
@dataclass
class SpecDecodingStats:
"""Per-step iteration decoding stats from scheduler.
Each scheduler step, statistics on spec decoding performance are
aggregated across requests by the scheduler and returned to the
frontend in EngineCoreOutputs->SchedulerStats.
"""
num_spec_tokens: int
num_drafts: int = 0
num_draft_tokens: int = 0
num_accepted_tokens: int = 0
num_accepted_tokens_per_pos: list[int] = field(default_factory=list)
@classmethod
def new(cls, num_spec_tokens: int) -> "SpecDecodingStats":
return cls(num_spec_tokens=num_spec_tokens,
num_accepted_tokens_per_pos=[0] * num_spec_tokens)
def observe_draft(self, num_draft_tokens: int, num_accepted_tokens: int):
self.num_drafts += 1
self.num_draft_tokens += num_draft_tokens
self.num_accepted_tokens += num_accepted_tokens
assert num_accepted_tokens <= self.num_spec_tokens
for i in range(num_accepted_tokens):
self.num_accepted_tokens_per_pos[i] += 1
class SpecDecodingLogging:
"""Aggregate and log spec decoding metrics.
LoggingStatLogger aggregates per-iteration metrics over a set
time interval using observe() and then logs them using log()
before resetting to zero.
"""
def __init__(self):
self.reset()
def reset(self):
self.num_drafts: list[int] = []
self.num_draft_tokens: list[int] = []
self.num_accepted_tokens: list[int] = []
self.accepted_tokens_per_pos_lists: list[list[int]] = []
def observe(self, spec_decoding_stats: SpecDecodingStats):
self.num_drafts.append(spec_decoding_stats.num_drafts)
self.num_draft_tokens.append(spec_decoding_stats.num_draft_tokens)
self.num_accepted_tokens.append(
spec_decoding_stats.num_accepted_tokens)
self.accepted_tokens_per_pos_lists.append(
spec_decoding_stats.num_accepted_tokens_per_pos)
def log(self, log_fn=logger.info):
if not self.num_drafts:
return
num_drafts = np.sum(self.num_drafts)
num_draft_tokens = np.sum(self.num_draft_tokens)
num_accepted_tokens = np.sum(self.num_accepted_tokens)
draft_acceptance_rate = (num_accepted_tokens / num_draft_tokens *
100 if num_draft_tokens > 0 else float("nan"))
# Conventionally, mean acceptance length includes the bonus token
mean_acceptance_length = 1 + (num_accepted_tokens / num_drafts)
pos_matrix = np.array(self.accepted_tokens_per_pos_lists)
acceptance_rates = np.sum(pos_matrix, axis=0) / num_drafts
rates_str = ", ".join(f"{p:.3f}" for p in acceptance_rates)
log_fn(
"SpecDecoding metrics: "
"Draft acceptance rate: %.1f%%, "
"Mean acceptance length: %.2f, "
"Accepted: %d tokens, "
"Drafted: %d tokens, "
"Per-position acceptance rate: %s",
draft_acceptance_rate,
mean_acceptance_length,
num_accepted_tokens,
num_draft_tokens,
rates_str,
)
self.reset()
class SpecDecodingProm:
"""Record spec decoding metrics in Prometheus.
The acceptance rate can be calculated using a PromQL query:
rate(vllm:spec_decode_num_accepted_tokens_total[$interval]) /
rate(vllm:spec_decode_num_draft_tokens_total[$interval])
The mean acceptance length (conventionally including bonus tokens)
can be calculated using:
1 + (
rate(vllm:spec_decode_num_accepted_tokens_total[$interval]) /
rate(vllm:spec_decode_num_drafts[$interval]))
A per-position acceptance rate vector can be computed using
vllm:spec_decode_num_accepted_tokens_per_pos[$interval] /
vllm:spec_decode_num_drafts[$interval]
"""
_counter_cls = prometheus_client.Counter
def __init__(
self,
speculative_config: Optional[SpeculativeConfig],
labelnames: list[str],
labelvalues: list[str],
):
self.spec_decoding_enabled = speculative_config is not None
if not self.spec_decoding_enabled:
return
self.counter_spec_decode_num_drafts = \
self._counter_cls(
name="vllm:spec_decode_num_drafts",
documentation="Number of spec decoding drafts.",
labelnames=labelnames).labels(*labelvalues)
self.counter_spec_decode_num_draft_tokens = \
self._counter_cls(
name="vllm:spec_decode_num_draft_tokens",
documentation="Number of draft tokens.",
labelnames=labelnames,).labels(*labelvalues)
self.counter_spec_decode_num_accepted_tokens = \
self._counter_cls(
name="vllm:spec_decode_num_accepted_tokens",
documentation="Number of accepted tokens.",
labelnames=labelnames).labels(*labelvalues)
assert speculative_config is not None
num_spec_tokens = (speculative_config.num_speculative_tokens
if self.spec_decoding_enabled else 0)
pos_labelnames = labelnames + ["position"]
base_counter = self._counter_cls(
name="vllm:spec_decode_num_accepted_tokens_per_pos",
documentation="Accepted tokens per draft position.",
labelnames=pos_labelnames,
)
self.counter_spec_decode_num_accepted_tokens_per_pos: list[
prometheus_client.Counter] = []
for pos in range(num_spec_tokens):
pos_labelvalues = labelvalues + [str(pos)]
self.counter_spec_decode_num_accepted_tokens_per_pos.append(
base_counter.labels(*pos_labelvalues))
def observe(self, spec_decoding_stats: SpecDecodingStats):
if not self.spec_decoding_enabled:
return
self.counter_spec_decode_num_drafts.inc(spec_decoding_stats.num_drafts)
self.counter_spec_decode_num_draft_tokens.inc(
spec_decoding_stats.num_draft_tokens)
self.counter_spec_decode_num_accepted_tokens.inc(
spec_decoding_stats.num_accepted_tokens)
for pos, counter in enumerate(
self.counter_spec_decode_num_accepted_tokens_per_pos):
counter.inc(spec_decoding_stats.num_accepted_tokens_per_pos[pos])

132
vllm/v1/spec_decode/ngram_proposer.py Normal file
View File

@@ -0,0 +1,132 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from typing import Optional
import numpy as np
from numba import jit
from vllm.config import VllmConfig
class NgramProposer:
def __init__(self, vllm_config: VllmConfig):
# Minimum length of the n-gram to match.
self.min_n = vllm_config.speculative_config.prompt_lookup_min
# Maximum length of the n-gram to match.
self.max_n = vllm_config.speculative_config.prompt_lookup_max
# Number of tokens follow the match. If there are less than k
# tokens follow the match, we will return the maximum amount of
# tokens until the end.
self.k = vllm_config.speculative_config.num_speculative_tokens
# Maximum length of the model.
self.max_model_len = vllm_config.model_config.max_model_len
# Trigger Numba JIT compilation for N-gram proposer.
# This usually takes less than 1 second.
self.propose(np.zeros(1024, dtype=np.int32))
def propose(
self,
context_token_ids: np.ndarray,
) -> Optional[np.ndarray]:
"""Proposes the next sequence of tokens based on n-gram pattern
matching in the context. The function finds matches of the last n
tokens in the previous context, and returns k tokens that followed
that match.
Args:
context_token_ids: Numpy array of token IDs representing the
context sequence.
Returns:
np.ndarray: The sequence of tokens that followed
the matched n-gram in the context.
None: If no matching n-gram pattern is found.
Example:
If context_token_ids = [1,2,3,4,2,3], min_n = 2, max_n = 3, and
k = 4:
- The last 3 (= max_n) tokens [4,2,3] cannot find a match.
- The last 2 tokens [2,3] will be matched against the previous
4 tokens [1,2,3,4].
- Finding a match of [2,3] would return the tokens that
followed that pattern. Here we will return [4,2,3] because
we only have three tokens after the match.
"""
# Do not generate draft tokens beyond the max model length.
k = min(self.k, self.max_model_len - context_token_ids.shape[0])
if k <= 0:
return None
# TODO(woosuk): Optimize this.
for n in range(self.max_n, self.min_n - 1, -1):
result = _find_subarray_kmp(context_token_ids, n, k)
if result is not None:
return result
return None
def load_model(self, *args, **kwargs):
# No model to load.
pass
@jit(nopython=True)
def _kmp_lps_array(pattern: np.ndarray) -> np.ndarray:
"""
Build the lps (longest proper prefix which is also suffix)
array for the pattern.
"""
lps = np.zeros(len(pattern), dtype=np.int32)
prev_lps = 0 # length of the previous longest prefix suffix
i = 1
while i < len(pattern):
if pattern[i] == pattern[prev_lps]:
prev_lps += 1
lps[i] = prev_lps
i += 1
else:
if prev_lps != 0:
prev_lps = lps[prev_lps - 1]
else:
lps[i] = 0
i += 1
return lps
@jit(nopython=True)
def _find_subarray_kmp(
context_token_ids: np.ndarray,
n: int,
k: int,
) -> Optional[np.ndarray]:
context_len = context_token_ids.shape[0]
assert n > 0
pattern = context_token_ids[-n:]
# Precompute lps array for Y
lps = _kmp_lps_array(pattern)
i = 0
j = 0
# -n because the last n tokens are used as pattern
while i < context_len - n:
if context_token_ids[i] == pattern[j]:
i += 1
j += 1
# If we have matched the entire Y
if j == n:
# Found pattern in context, gather the next K elements
return context_token_ids[i:i + k]
else:
# Mismatch
if j != 0:
# Use the lps array to avoid re-checking elements
j = lps[j - 1]
else:
i += 1
# Y not found
return None

46
vllm/v1/spec_decode/utils.py Normal file
View File

@@ -0,0 +1,46 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from vllm.triton_utils import tl, triton
from vllm.v1.worker.gpu_input_batch import InputBatch
def is_spec_decode_supported(req_id: str, input_batch: InputBatch) -> bool:
if req_id in input_batch.min_p_reqs:
# Spec decode doesn't support min_p sampling.
return False
elif (req_id in input_batch.frequency_penalties_reqs
or req_id in input_batch.presence_penalties_reqs
or req_id in input_batch.repetition_penalties_reqs):
# Spec decode doesn't support penalties.
return False
elif req_id in input_batch.num_logprobs:
# Spec decode doesn't support logprobs.
return False
return True
@triton.jit
def prepare_eagle_input_kernel(
out_ptr,
cu_query_lens_ptr,
cu_num_tokens_ptr,
BLOCK_SIZE: tl.constexpr,
):
pid = tl.program_id(0)
# [start_pos, end_pos)
start_pos = tl.load(cu_num_tokens_ptr + pid)
end_pos = tl.load(cu_num_tokens_ptr + pid + 1)
num_tokens = end_pos - start_pos
index_start = tl.load(cu_query_lens_ptr + pid)
num_blocks = tl.cdiv(num_tokens, BLOCK_SIZE)
for i in tl.range(num_blocks):
offset = i * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
tl.store(
out_ptr + start_pos + offset,
index_start + offset,
mask=offset < num_tokens,
)