class Sampler:
def __init__(
self,
max_num_reqs: int,
vocab_size: int,
device: torch.device,
req_states: RequestState,
logprobs_mode: LogprobsMode = "raw_logprobs",
num_speculative_tokens: int = 1,
use_fp64_gumbel: bool = False,
):
self.logprobs_mode = logprobs_mode
self.compute_nans = envs.VLLM_COMPUTE_NANS_IN_LOGITS # False by default.
self.use_fp64_gumbel = use_fp64_gumbel
self.req_states = req_states
self.sampling_states = SamplingStates(max_num_reqs, vocab_size)
self.penalties_state = PenaltiesState(req_states)
self.logit_bias_state = LogitBiasState(max_num_reqs, device)
self.bad_words_state = BadWordsState(req_states)
self.logprob_token_ids_state = LogprobTokenIdsState(max_num_reqs, device)
self.num_speculative_tokens = num_speculative_tokens
self.use_flashinfer = flashinfer_sampler_supported()
def add_request(
self, req_idx: int, prompt_len: int, sampling_params: SamplingParams
) -> None:
self.sampling_states.add_request(req_idx, sampling_params)
self.penalties_state.add_request(req_idx, sampling_params)
self.logit_bias_state.add_request(req_idx, prompt_len, sampling_params)
self.bad_words_state.add_request(req_idx, sampling_params)
self.logprob_token_ids_state.add_request(req_idx, sampling_params)
def apply_staged_writes(self) -> None:
self.sampling_states.apply_staged_writes()
self.penalties_state.apply_staged_writes()
self.logit_bias_state.apply_staged_writes()
self.bad_words_state.apply_staged_writes()
self.logprob_token_ids_state.apply_staged_writes()
def __call__(
self,
logits: torch.Tensor,
input_batch: InputBatch,
) -> SamplerOutput:
expanded_idx_mapping = input_batch.expanded_idx_mapping
idx_mapping_np = input_batch.idx_mapping_np
cu_num_logits_np = input_batch.cu_num_logits_np
expanded_local_pos = input_batch.expanded_local_pos
pos = input_batch.positions[input_batch.logits_indices]
input_ids = input_batch.input_ids[input_batch.logits_indices]
# NOTE(woosuk): We intentionally compute num_nans before sampling to make clear
# that num_nans is computed before applying penalties and temperature.
num_nans = get_num_nans(logits) if self.compute_nans else None
return_logprobs = self.returns_logprobs(idx_mapping_np)
sampled, processed_logits = self.sample(
logits,
expanded_idx_mapping,
idx_mapping_np,
pos,
input_ids,
expanded_local_pos,
return_logprobs=return_logprobs,
)
if return_logprobs:
if self.logprobs_mode in ("processed_logprobs", "processed_logits"):
logits = processed_logits
max_num_logprobs = self.sampling_states.max_num_logprobs(idx_mapping_np)
max_per_req_token_ids = self.logprob_token_ids_state.max_num_token_ids(
idx_mapping_np
)
expanded_logits = logits.shape[0] != idx_mapping_np.shape[0]
cu_num_logits = cu_num_logits_np.tolist() if expanded_logits else None
num_logprobs = max_num_logprobs if max_num_logprobs != NO_LOGPROBS else 0
logprobs_tensors = compute_topk_scores(
logits,
num_logprobs,
sampled,
cu_num_logits,
logprob_token_ids_state=self.logprob_token_ids_state,
expanded_idx_mapping=input_batch.expanded_idx_mapping,
max_per_req_token_ids=max_per_req_token_ids,
logits_mode=self.logprobs_mode in ("raw_logits", "processed_logits"),
)
else:
logprobs_tensors = None
# 1 sampled token per request, except chunked-prefill requests
# (seq_len < prefill_len) which aren't done prefilling and produce no
# output token. num_rejected is always 0 here (one logit per request).
num_sampled, num_rejected = get_num_sampled_and_rejected(
input_batch.seq_lens.new_ones(input_batch.num_reqs),
input_batch.seq_lens,
input_batch.cu_num_logits,
input_batch.idx_mapping,
self.req_states.prefill_len.gpu,
)
# These are GPU tensors.
sampler_output = SamplerOutput(
# The sampled tokens are expanded to 2D tensor with shape
# [num_requests, 1], where each row represents one generated
# token per request.
sampled_token_ids=sampled.view(-1, 1),
logprobs_tensors=logprobs_tensors,
num_nans=num_nans,
num_sampled=num_sampled,
num_rejected=num_rejected,
)
return sampler_output
def returns_logprobs(self, idx_mapping_np: np.ndarray) -> bool:
"""Whether any request in the batch produces logprobs this step."""
return (
self.sampling_states.max_num_logprobs(idx_mapping_np) != NO_LOGPROBS
or self.logprob_token_ids_state.max_num_token_ids(idx_mapping_np) > 0
)
def apply_sampling_params(
self,
logits: torch.Tensor,
expanded_idx_mapping: torch.Tensor,
idx_mapping_np: np.ndarray,
pos: torch.Tensor,
input_ids: torch.Tensor,
expanded_local_pos: torch.Tensor,
skip_top_k_top_p: bool = False,
) -> torch.Tensor:
# The ops below upcast to fp32 internally, so the input dtype is kept and
# mutated in place. Only raw_logprobs reads the unmodified logits
# afterward, so copy just for that case.
if self.logprobs_mode.startswith("raw_") and self.returns_logprobs(
idx_mapping_np
):
logits = logits.clone()
# Apply logit bias (e.g., allowed_token_ids, min_tokens) in place.
self.logit_bias_state.apply_logit_bias(
logits, expanded_idx_mapping, idx_mapping_np, pos
)
# Apply penalties in place.
self.penalties_state.apply_penalties(
logits,
expanded_idx_mapping,
idx_mapping_np,
input_ids,
expanded_local_pos,
)
# Apply bad words masking in place.
self.bad_words_state.apply_bad_words(
logits,
expanded_idx_mapping,
idx_mapping_np,
input_ids,
expanded_local_pos,
)
# Apply temperature in place.
self.sampling_states.apply_temperature(
logits, expanded_idx_mapping, idx_mapping_np
)
# Apply min_p in place.
self.sampling_states.apply_min_p(logits, expanded_idx_mapping, idx_mapping_np)
if skip_top_k_top_p:
return logits
# Apply top_k and/or top_p. This might or might not return a new tensor.
return self.sampling_states.apply_top_k_top_p(
logits, expanded_idx_mapping, idx_mapping_np
)
def sample(
self,
logits: torch.Tensor,
expanded_idx_mapping: torch.Tensor,
idx_mapping_np: np.ndarray,
pos: torch.Tensor,
input_ids: torch.Tensor,
expanded_local_pos: torch.Tensor,
return_logprobs: bool = False,
) -> tuple[torch.Tensor, torch.Tensor]:
processed_logits = self.apply_sampling_params(
logits,
expanded_idx_mapping,
idx_mapping_np,
pos,
input_ids,
expanded_local_pos,
skip_top_k_top_p=True,
)
top_k, top_p = self.sampling_states.get_top_k_top_p(
expanded_idx_mapping, idx_mapping_np
)
use_flashinfer = self.use_flashinfer and not (
# Don't use FI sampler if no requests use top_k/top_p, if there are
# any greedy requests or per-request seeds, or if post-processed
# logprobs need to be returned for any requests.
(top_k is None and top_p is None)
or (
return_logprobs
and self.logprobs_mode in ("processed_logprobs", "processed_logits")
)
or self.sampling_states.any_greedy(idx_mapping_np)
or self.sampling_states.any_explicit_seed(idx_mapping_np)
)
# Sample the next token.
if use_flashinfer:
sampled = flashinfer_sample(processed_logits, top_k, top_p).to(torch.int64)
else:
processed_logits = apply_top_k_top_p(processed_logits, top_k, top_p)
sampled = gumbel_sample(
processed_logits,
expanded_idx_mapping,
self.sampling_states.temperature.gpu,
self.sampling_states.seeds.gpu,
pos,
apply_temperature=False,
use_fp64=self.use_fp64_gumbel,
)
return sampled, processed_logits