[V1] LogitsProcessor programming model (#16728)

Signed-off-by: Nick Hill <nhill@redhat.com>
Signed-off-by: Andrew Feldman <afeldman@neuralmagic.com>
Signed-off-by: Andrew Feldman <afeldman@redhat.com>
Co-authored-by: Nick Hill <nhill@redhat.com>
This commit is contained in:
afeldman-nm
2025-07-02 09:10:42 -07:00
committed by GitHub
co-authored by Nick Hill
parent c1909e7e8c
commit 48fb076cbc
13 changed files with 1401 additions and 393 deletions
+11 -71
View File
@@ -5,12 +5,11 @@
import torch
import torch.nn as nn
from vllm.utils import async_tensor_h2d, is_pin_memory_available
from vllm.utils import is_pin_memory_available
from vllm.v1.outputs import LogprobsTensors, SamplerOutput
from vllm.v1.sample.metadata import SamplingMetadata
from vllm.v1.sample.ops.bad_words import apply_bad_words
from vllm.v1.sample.ops.penalties import (apply_all_penalties,
apply_min_token_penalties)
from vllm.v1.sample.ops.penalties import apply_all_penalties
from vllm.v1.sample.ops.topk_topp_sampler import TopKTopPSampler
_SAMPLING_EPS = 1e-5
@@ -44,8 +43,11 @@ class Sampler(nn.Module):
logits = self.apply_allowed_token_ids(logits, sampling_metadata)
# Apply bad words exclusion.
logits = self.apply_bad_words(logits, sampling_metadata)
# Apply logits bias.
logits = self.apply_logits_bias(logits, sampling_metadata)
# Apply logits processors which can impact greedy sampling
for processor in (sampling_metadata.logitsprocs.non_argmax_invariant):
logits = processor.apply(logits)
# Apply penalties (e.g., min_tokens, freq_penalties).
logits = self.apply_penalties(logits, sampling_metadata)
# Sample the next token.
@@ -110,9 +112,10 @@ class Sampler(nn.Module):
# Apply temperature.
logits = self.apply_temperature(logits, sampling_metadata.temperature)
# Apply min_p.
if sampling_metadata.min_p is not None:
logits = self.apply_min_p(logits, sampling_metadata.min_p)
# Apply logits processors that only apply to random sampling
# (argmax invariant)
for processor in sampling_metadata.logitsprocs.argmax_invariant:
logits = processor.apply(logits)
# Apply top_k and/or top_p.
random_sampled = self.topk_topp_sampler(
@@ -187,10 +190,6 @@ class Sampler(nn.Module):
logits: torch.Tensor,
sampling_metadata: SamplingMetadata,
) -> torch.Tensor:
if sampling_metadata.min_tokens:
apply_min_token_penalties(logits,
sampling_metadata.output_token_ids,
sampling_metadata.min_tokens)
if not sampling_metadata.no_penalties:
assert sampling_metadata.prompt_token_ids is not None
logits = apply_all_penalties(
@@ -203,65 +202,6 @@ class Sampler(nn.Module):
)
return logits
def apply_min_p(
self,
logits: torch.Tensor,
min_p: torch.Tensor,
) -> torch.Tensor:
"""
Filters logits using adaptive probability thresholding.
"""
# Convert logits to probability distribution
probability_values = torch.nn.functional.softmax(logits, dim=-1)
# Calculate maximum probabilities per sequence
max_probabilities = torch.amax(probability_values,
dim=-1,
keepdim=True)
# Reshape min_p for broadcasting
adjusted_min_p = min_p.unsqueeze(1) * max_probabilities
# Identify valid tokens using threshold comparison
valid_token_mask = probability_values >= adjusted_min_p
# Apply mask using boolean indexing
logits[~valid_token_mask] = -float('inf')
return logits
def apply_logits_bias(
self,
logits: torch.Tensor,
sampling_metadata: SamplingMetadata,
) -> torch.Tensor:
# TODO(houseroad): this implementation is extremely inefficient.
# One idea is implement this as a PyTorch C++ op, and we may
# even optimize the logit_bias layout.
rows: list[int] = []
cols: list[int] = []
vals: list[float] = []
# Get vocabulary size from logits
vocab_size = logits.shape[-1]
for i, logit_bias in enumerate(sampling_metadata.logit_bias):
if logit_bias:
for token_id, bias in logit_bias.items():
# Check token_id bounds to ensure within vocabulary
if token_id < 0 or token_id >= vocab_size:
raise ValueError(
f"token_id {token_id} in logit_bias contains "
f"out-of-vocab token id. Vocabulary size: "
f"{vocab_size}")
rows.append(i)
cols.append(token_id)
vals.append(bias)
if rows:
indices = async_tensor_h2d([rows, cols], torch.int64,
logits.device, self.pin_memory)
values = async_tensor_h2d(vals, torch.float, logits.device,
self.pin_memory)
logits.index_put_(tuple(indices), values=values, accumulate=True)
return logits
def apply_allowed_token_ids(
self,
logits: torch.Tensor,