refactor: split infer core into subpackages by concern

- Eliminate core/ directory into cache/, runtime/, network/ subpackages plus flat modules
- Split cache.py (647 lines) into cache/{buffer,strategy,pool}.py by layer
- Add explicit ContiguousStrategy, make AllocationStrategy a real ABC
- Move TaskCacheState to cache/strategy.py, drop string forward references
- Rename api/ to network/, server.py to app.py
- Move sample.py into runtime/ alongside executor and graph
- Simplify TaskCacheManager.__init__ to single pool param
- Expose pool.strategy and pool.req_pool as public properties
- Fix KVCache import in attention_backend.py (TYPE_CHECKING guard)
- Fix steady-state decode reading uninitialized position_ids on first step
This commit is contained in:
2026-08-08 23:43:05 +08:00
parent 3fa7e66676
commit 0c1b7664c1
37 changed files with 920 additions and 800 deletions
+25
View File
@@ -0,0 +1,25 @@
"""Execution primitives: forward passes, CUDA graphs, and sampling."""
from astrai.inference.runtime.executor import Executor
from astrai.inference.runtime.graph import CudaGraphContext
from astrai.inference.runtime.sample import (
BaseSamplingStrategy,
FrequencyPenaltyStrategy,
SamplingPipeline,
TemperatureStrategy,
TopKStrategy,
TopPStrategy,
sample,
)
__all__ = [
"Executor",
"CudaGraphContext",
"BaseSamplingStrategy",
"FrequencyPenaltyStrategy",
"SamplingPipeline",
"TemperatureStrategy",
"TopKStrategy",
"TopPStrategy",
"sample",
]
+418
View File
@@ -0,0 +1,418 @@
import logging
import time
from contextlib import contextmanager
from dataclasses import dataclass
from typing import List, Optional
import torch
from torch import Tensor
from astrai.extension.attention_backend import (
ATTN_BACKEND,
CudaBackend,
attn_backend,
get_backend,
)
from astrai.inference.cache import PagePool, TaskCacheManager
from astrai.inference.runtime.graph import CudaGraphContext
from astrai.inference.runtime.sample import sample
from astrai.inference.task import Task
from astrai.inference.workspace import InferenceWorkspace
from astrai.model.automodel import AutoModel
logger = logging.getLogger(__name__)
@contextmanager
def timed(label: str, log: Optional[logging.Logger] = None):
"""GPU-precise timer via CUDA events; falls back to perf_counter on CPU."""
log = log or logger
if not log.isEnabledFor(logging.DEBUG):
yield
return
use_cuda = torch.cuda.is_available()
if use_cuda:
start = torch.cuda.Event(enable_timing=True)
end = torch.cuda.Event(enable_timing=True)
start.record()
else:
tic = time.perf_counter()
yield
if use_cuda:
end.record()
torch.cuda.synchronize()
elapsed_ms = start.elapsed_time(end)
else:
elapsed_ms = (time.perf_counter() - tic) * 1000
log.debug("%s %.2fms", label, elapsed_ms)
@dataclass
class SamplingBatchInfo:
"""Per-batch sampling parameters, cached across decode steps.
Sampling params are constant for a given ordered task set, so they are
built once (pinned-memory async H2D) and reused until the task set
changes. ``top_ks`` is int32 to match the native consumers.
"""
temperatures: Tensor # float32 [B]
top_ks: Tensor # int32 [B]
top_ps: Tensor # float32 [B]
freq_penalties: Tensor # float32 [B]
has_freq: bool # any frequency_penalty != 0 (avoids per-step GPU .any())
@dataclass
class DecodeSteadyState:
"""Cached decode metadata for the steady-state case.
When the same ordered task set decodes one token per step, sampling
params and task signature are reused; only positions advance by 1.
"""
task_sig: tuple
positions: list[int]
sampling_info: SamplingBatchInfo
def _build_sampling_batch_info(tasks: List[Task], device) -> SamplingBatchInfo:
pin = str(device).startswith("cuda")
freq_penalties = torch.tensor(
[t.frequency_penalty for t in tasks], dtype=torch.float32, pin_memory=pin
).to(device, non_blocking=True)
return SamplingBatchInfo(
temperatures=torch.tensor(
[t.temperature for t in tasks], dtype=torch.float32, pin_memory=pin
).to(device, non_blocking=True),
top_ks=torch.tensor(
[t.top_k for t in tasks], dtype=torch.int32, pin_memory=pin
).to(device, non_blocking=True),
top_ps=torch.tensor(
[t.top_p for t in tasks], dtype=torch.float32, pin_memory=pin
).to(device, non_blocking=True),
freq_penalties=freq_penalties,
has_freq=bool((freq_penalties != 0).any()),
)
def _warmup_cuda_graphs(
model: AutoModel,
pool: PagePool,
task_cache: TaskCacheManager,
ws: InferenceWorkspace,
gctx: CudaGraphContext,
max_batch_size: int,
prompt_len: int = 1,
device: Optional[str] = None,
):
dev = device or next(model.parameters()).device
# Prefill warmup: cuBLAS auto-tunes for the actual prompt-length tensor
# shapes on first call (F.linear is the dominant cost). This also warms
# up the CUDA context (driver init) and compiles the graph-capture trace
# that follows. Custom .so kernels do NOT need this — they are pre-built.
warmup_len = 64
tid = "_warmup_prefill"
if task_cache.task_alloc(tid, list(range(warmup_len))):
with (
torch.inference_mode(),
timed("warmup prefill", logger),
):
kv = task_cache.bind([tid], ws, start_pos=0)
ids_in = torch.arange(warmup_len, device=dev).unsqueeze(0)
pos_in = ids_in
model(
ids_in,
input_mask=pos_in.unsqueeze(-1) >= torch.arange(warmup_len, device=dev),
kv_cache=kv,
position_ids=pos_in,
)
task_cache.task_free(tid)
batch_sizes = [1]
n = 2
while n <= max_batch_size:
batch_sizes.append(n)
n *= 2
if max_batch_size not in batch_sizes:
batch_sizes.append(max_batch_size)
for b in batch_sizes:
task_ids = [f"_warmup_decode_{b}_{i}" for i in range(b)]
prompt_tokens = [list(range(prompt_len)) for _ in range(b)]
alloc_ok = True
for tid, pt in zip(task_ids, prompt_tokens):
if not task_cache.task_alloc(tid, pt):
alloc_ok = False
break
if not alloc_ok:
for tid in task_ids:
task_cache.task_free(tid)
continue
with (
torch.inference_mode(),
attn_backend(ATTN_BACKEND.CUDA),
timed(f"warmup decode b={b}", logger),
):
for step in range(2):
seq_pos = step
ws.position_ids[:b] = seq_pos
for tid in task_ids:
task_cache.task_extend(tid, seq_pos)
kv = task_cache.bind(task_ids, ws)
input_mask = ws.decode_mask(ws.position_ids[:b], ws.max_seq_len)
ids_buf = ws.fill_input_ids([step] * b)
gctx.forward(
model,
key=(b,),
input_ids=ids_buf.unsqueeze(1),
input_mask=input_mask,
kv_cache=kv,
position_ids=ws.position_ids[:b].unsqueeze(1),
)
for tid in task_ids:
task_cache.task_free(tid)
torch.cuda.synchronize()
class Executor:
"""Model forward passes for prefill and decode phases."""
def __init__(
self,
model: AutoModel,
kv_cache: PagePool,
task_cache: TaskCacheManager,
device: Optional[str] = None,
dtype: Optional[torch.dtype] = None,
):
self.model = model
self.kv_cache = kv_cache
self.task_cache = task_cache
self.device = device or next(model.parameters()).device
self.dtype = dtype or next(model.parameters()).dtype
# Per-step decode cache for the steady-state case (same ordered
# task set decodes one token per step). Sampling params stay
# constant; only positions advance.
self._decode_cache: Optional[DecodeSteadyState] = None
# Pre-allocated fixed-shape buffers for the decode hot path
# (input_ids, decode mask, KV bind metadata). Eagerly sized at init
# so the workspace is CUDA-graph-capture friendly — no allocation
# during capture.
config = model.config
max_q_heads = config.num_attention_heads
head_dim = config.hidden_size // config.num_attention_heads
self._graph_supported = CudaBackend.supports(head_dim=head_dim)
self._workspace = InferenceWorkspace(
max_batch_size=kv_cache.max_batch_size,
max_seq_len=kv_cache.max_seq_len,
max_q_heads=max_q_heads,
head_dim=head_dim,
device=self.device,
dtype=self.dtype,
)
# CUDA-graph capture: one graph per (batch_size,) key.
# Enabled at init-time via _warmup_cuda_graphs for CudaBackend
# on supported head_dims; left disabled otherwise.
self._graph_ctx = CudaGraphContext()
self._try_enable_cuda_graph()
def _try_enable_cuda_graph(self):
if not self._graph_supported:
return
self._graph_ctx.set_enabled(True)
_warmup_cuda_graphs(
self.model,
self.kv_cache,
self.task_cache,
self._workspace,
self._graph_ctx,
max_batch_size=self.kv_cache.max_batch_size,
device=self.device,
)
def _sample_logits(
self,
logits: Tensor,
tasks: List[Task],
return_logprobs: bool = False,
info: Optional[SamplingBatchInfo] = None,
):
info = info or _build_sampling_batch_info(tasks, self.device)
if info.has_freq:
history_lists = [
t.prompt_ids[-t.rep_window :] + t.output_ids for t in tasks
]
history_lens = [len(ids) for ids in history_lists]
max_len = max(history_lens, default=0)
padded_ids = torch.zeros(
len(tasks), max_len, dtype=torch.long, device=self.device
)
padded_mask = torch.zeros(
len(tasks), max_len, dtype=torch.bool, device=self.device
)
for i, ids in enumerate(history_lists):
length = len(ids)
padded_ids[i, :length] = torch.as_tensor(
ids, dtype=torch.long, device=self.device
)
padded_mask[i, :length] = True
else:
padded_ids = None
padded_mask = None
result = sample(
logits,
temperature=info.temperatures,
top_k=info.top_ks,
top_p=info.top_ps,
frequency_penalty=info.freq_penalties,
input_ids=padded_ids,
input_mask=padded_mask,
return_logprobs=return_logprobs,
)
if not return_logprobs:
return result.tolist()
tokens, logprobs = result
tokens_list = tokens.tolist()
logprobs_list = logprobs.tolist()
for task, logprob in zip(tasks, logprobs_list):
task.output_logprobs.append(float(logprob))
return list(zip(tokens_list, logprobs_list))
def execute_prefill(
self,
tasks: List[Task],
prompt_len: int,
start_pos: int = 0,
return_logprobs: bool = False,
):
if start_pos >= prompt_len:
return []
tasks = sorted(tasks, key=lambda t: t.task_id)
batch_sz = len(tasks)
input_ids = torch.tensor(
[t.prompt_ids[start_pos:prompt_len] for t in tasks],
dtype=torch.long,
device=self.device,
)
task_ids = [t.task_id for t in tasks]
position_ids = (
torch.arange(start_pos, prompt_len, dtype=torch.long, device=self.device)
.unsqueeze(0)
.expand(batch_sz, -1)
)
input_mask = position_ids.unsqueeze(-1) >= torch.arange(
prompt_len, device=self.device
)
with (
torch.inference_mode(),
timed(f"execute_prefill b={batch_sz} prompt_len={prompt_len}", logger),
):
outputs = self.model(
input_ids,
input_mask=input_mask,
position_ids=position_ids,
kv_cache=self.task_cache.bind(
task_ids,
self._workspace,
start_pos=start_pos,
),
)
logits = outputs["logits"][:, -1, :]
return tasks, self._sample_logits(logits, tasks, return_logprobs)
def execute_decode(
self, tasks: List[Task], return_logprobs: bool = False
) -> List[int]:
"""Decode next token for each task.
Args:
return_logprobs: When ``True``, also record (and return)
the log-probability of each sampled token under the
post-strategy sampling distribution. The logprob is
appended to ``task.output_logprobs`` and the return
list becomes ``List[Tuple[int, float]]``.
Returns:
``List[int]`` of sampled token IDs, or
``List[Tuple[int, float]]`` of ``(token_id, logprob)`` when
``return_logprobs`` is ``True``.
"""
if not tasks:
return []
b = len(tasks)
ws = self._workspace
# ---- pre-replay: update input buffers in-place ----
input_ids = ws.fill_input_ids(
[t.output_ids[-1] if t.output_ids else t.prompt_ids[-1] for t in tasks]
)
task_ids = [t.task_id for t in tasks]
cur_positions = [t.next_pos for t in tasks]
kv_cache = self.task_cache.bind(task_ids, ws)
if self.task_cache.bind_was_steady and self._decode_cache is not None:
info = self._decode_cache.sampling_info
ws.position_ids[:b] += 1
else:
info = _build_sampling_batch_info(tasks, self.device)
ws.position_ids[:b].copy_(
torch.tensor(cur_positions, dtype=torch.long, device=self.device)
)
self._decode_cache = DecodeSteadyState(tuple(task_ids), cur_positions, info)
total_len = max(cur_positions) + 1
input_mask = ws.decode_mask(ws.position_ids[:b], total_len)
# ---- forward (graph replay or live run + capture) ----
use_graph = (
self._graph_ctx.enabled
and self._graph_supported
and get_backend().supports_graph()
)
key = (b,)
if use_graph:
input_mask = ws.decode_mask(ws.position_ids[:b], ws.max_seq_len)
with (
torch.inference_mode(),
timed(f"execute_decode forward b={b}", logger),
):
if use_graph:
outputs = self._graph_ctx.forward(
self.model,
key=key,
input_ids=input_ids.unsqueeze(1),
input_mask=input_mask,
kv_cache=kv_cache,
position_ids=ws.position_ids[:b].unsqueeze(1),
)
else:
outputs = self.model(
input_ids.unsqueeze(1),
input_mask=input_mask,
kv_cache=kv_cache,
position_ids=ws.position_ids[:b].unsqueeze(1),
)
logits = outputs["logits"][:, -1, :]
return self._sample_logits(logits, tasks, return_logprobs, info=info)
+103
View File
@@ -0,0 +1,103 @@
"""CUDA-graph capture for the decode model-forward step.
Mirrors SGLang's cuda-graph manager: one graph per batch size. The graph
pair. The graph captures ``model.forward()`` with workspace-backed inputs
(all at fixed addresses). Before each replay the caller updates the input
buffer content in-place so the graph sees fresh data at the same tensor
addresses.
Only the model forward is captured — sampling runs outside the graph
(via ``torch.multinomial`` which consumes a mutable RNG state).
"""
import torch
from torch import Tensor
class CudaGraphContext:
"""CUDA-graph capture/replay for decode steps.
Parameters:
enabled: When ``False``, ``forward()`` always runs the live model
forward without capture/replay (graphs are cleared). Toggle at
runtime via the ``set_enabled()`` method.
Usage::
gctx = CudaGraphContext()
with torch.inference_mode():
outputs = gctx.forward(
model,
key=(batch_size,),
input_ids=workspace.input_ids[:b].unsqueeze(1),
input_mask=input_mask,
kv_cache=kv_cache,
position_ids=workspace.position_ids[:b].unsqueeze(1),
)
The first call at a given key runs *without* capture (warmup). The
second call captures the graph. Subsequent calls replay the captured
graph. A ``torch.cuda.synchronize()`` before capture drains in-flight
work so the graph trace is clean.
"""
def __init__(self, enabled: bool = False):
self._enabled = enabled
self._graphs: dict[tuple, torch.cuda.CUDAGraph] = {}
self._outputs: dict[tuple, dict[str, Tensor]] = {}
self._warmed: set[tuple] = set()
@property
def enabled(self) -> bool:
return self._enabled
def set_enabled(self, flag: bool):
"""Enable or disable CUDA-graph capture at runtime.
Disabling clears all captured graphs (frees GPU memory) and warmup
state. Re-enabling after disable starts fresh — graphs are
re-captured on the next warmup cycle.
"""
if flag == self._enabled:
return
self._enabled = flag
if not flag:
self._graphs.clear()
self._outputs.clear()
self._warmed.clear()
def forward(self, model, *, key, **kwargs) -> dict[str, Tensor]:
"""Run ``model(**kwargs)`` via graph replay or live forward.
Args:
model: callable, e.g. ``self.model.forward``.
key: ``(batch_size,)`` — the dispatch key (one graph per batch size).
**kwargs: arguments forwarded to ``model``. All tensor arguments
must reside at stable addresses (workspace buffers).
Returns:
The dict produced by ``model(**kwargs)``, e.g.
``{"logits": ..., "h0": ...}``.
"""
if not self._enabled:
self._outputs[key] = model(**kwargs)
return self._outputs[key]
if key in self._graphs:
self._graphs[key].replay()
elif key in self._warmed:
cap_output = model(**kwargs)
torch.cuda.synchronize()
graph = torch.cuda.CUDAGraph()
with torch.cuda.graph(graph):
self._outputs[key] = model(**kwargs)
self._graphs[key] = graph
self._warmed.discard(key)
return cap_output
else:
self._warmed.add(key)
self._outputs[key] = model(**kwargs)
return self._outputs[key]
def has_graph(self, key: tuple) -> bool:
return key in self._graphs
+386
View File
@@ -0,0 +1,386 @@
"""Composable sampling strategies for logit transformation.
Implements the Strategy pattern: each sampling technique
(temperature, top-k, top-p, frequency penalty) is a pluggable
strategy that can be composed into a pipeline.
All strategies accept both scalar and per-sample tensor
parameters, so a single pipeline works for any batch size.
"""
from abc import ABC, abstractmethod
from typing import List, Optional, Union
import torch
from torch import Tensor
class BaseSamplingStrategy(ABC):
"""Abstract base for a logit transformation strategy."""
@abstractmethod
def apply(
self,
logits: Tensor,
filter_value: float = -float("inf"),
input_ids: Optional[Tensor] = None,
input_mask: Optional[Tensor] = None,
) -> Tensor:
"""Applies the strategy to logits.
Args:
logits: Raw logits tensor (batch, vocab_size).
filter_value: Value assigned to filtered-out positions.
input_ids: Previously generated token IDs ``[batch, seq_len]``,
padded with 0. Used by frequency penalty.
input_mask: Boolean mask ``[batch, seq_len]``, True for real
tokens, False for padding. Used to exclude padding from
penalty computation.
Returns:
Transformed logits tensor.
"""
raise NotImplementedError
class TemperatureStrategy(BaseSamplingStrategy):
"""Divides logits by temperature to control randomness.
Args:
temperature: Scalar or ``[batch]`` tensor.
"""
def __init__(self, temperature: Union[float, Tensor] = 1.0):
self.temperature = temperature
def apply(
self,
logits: Tensor,
filter_value: float = -float("inf"),
input_ids: Optional[Tensor] = None,
input_mask: Optional[Tensor] = None,
) -> Tensor:
t = self.temperature
if isinstance(t, Tensor):
t = t.to(logits.device, non_blocking=True).view(-1, 1)
t = torch.clamp(t, min=1e-8)
if (t != 1.0).any():
logits = logits / t
elif t != 1.0:
logits = logits / max(t, 1e-8)
return logits
class TopKStrategy(BaseSamplingStrategy):
"""Keeps only the top-k logits, setting the rest to filter_value.
Args:
top_k: Scalar or ``[batch]`` tensor (0 disables).
"""
def __init__(self, top_k: Union[int, Tensor] = 0):
self.top_k = top_k
def apply(
self,
logits: Tensor,
filter_value: float = -float("inf"),
input_ids: Optional[Tensor] = None,
input_mask: Optional[Tensor] = None,
) -> Tensor:
tk = self.top_k
if isinstance(tk, Tensor):
tk = tk.to(logits.device, non_blocking=True).long().clamp(min=0)
max_k = int(tk.max().item())
if max_k <= 0:
return logits
max_k = min(max_k, logits.size(-1))
values, _ = torch.topk(logits, max_k, dim=-1)
per_row_k = tk.clamp(max=max_k)
thresholds = torch.full_like(logits[..., -1:], -float("inf"))
positive = per_row_k > 0
if positive.any():
row_idx = torch.arange(logits.size(0), device=logits.device)[positive]
thresholds[positive] = values[
row_idx, per_row_k[positive] - 1
].unsqueeze(-1)
logits[logits < thresholds] = filter_value
return logits
if tk > 0:
k = min(tk, logits.size(-1))
thresholds = torch.topk(logits, k, dim=-1)[0][..., -1:]
logits[logits < thresholds] = filter_value
return logits
class TopPStrategy(BaseSamplingStrategy):
"""Nucleus (top-p) filtering: keeps the smallest set of tokens whose
cumulative probability exceeds top_p.
Args:
top_p: Scalar or ``[batch]`` tensor (1.0 disables).
"""
def __init__(self, top_p: Union[float, Tensor] = 1.0):
self.top_p = top_p
def _apply(
self, logits: Tensor, top_p: Union[float, Tensor], filter_value: float
) -> Tensor:
sorted_logits, sorted_indices = torch.sort(logits, descending=True, dim=-1)
cum_probs = torch.cumsum(torch.softmax(sorted_logits, dim=-1), dim=-1)
remove = cum_probs > top_p
remove[..., 1:] = remove[..., :-1].clone()
remove[..., 0] = False
mask = torch.zeros_like(logits, dtype=torch.bool)
mask.scatter_(1, sorted_indices, remove)
logits[mask] = filter_value
return logits
def apply(
self,
logits: Tensor,
filter_value: float = -float("inf"),
input_ids: Optional[Tensor] = None,
input_mask: Optional[Tensor] = None,
) -> Tensor:
tp = self.top_p
if isinstance(tp, Tensor):
tp = tp.to(logits.device, non_blocking=True)
if (tp < 1.0).any():
logits = self._apply(logits, tp.view(-1, 1), filter_value)
elif tp < 1.0:
logits = self._apply(logits, tp, filter_value)
return logits
class FrequencyPenaltyStrategy(BaseSamplingStrategy):
"""Penalizes tokens based on how many times they appeared in history.
Subtracts ``penalty * count(token)`` from each token's logit, where
``count(token)`` is the number of occurrences in the generation history
(prompt + output). A penalty of ``0.0`` disables the strategy.
Unlike repetition penalty (which only checks *presence*), frequency
penalty scales linearly with occurrence count: the first use is
penalized once, the third use three times. This allows natural
repetition of common words while suppressing degenerate loops.
Reference: OpenAI API ``frequency_penalty`` parameter.
Args:
penalty: Scalar or ``[batch]`` tensor (0.0 disables, range -2.0~2.0).
"""
def __init__(self, penalty: Union[float, Tensor] = 0.0):
self.penalty = penalty
def apply(
self,
logits: Tensor,
filter_value: float = -float("inf"),
input_ids: Optional[Tensor] = None,
input_mask: Optional[Tensor] = None,
) -> Tensor:
if input_ids is None:
return logits
p = self.penalty
if isinstance(p, Tensor):
p = p.to(logits.device, non_blocking=True).view(-1, 1)
if (p == 0.0).all():
return logits
elif p == 0.0:
return logits
input_ids = input_ids.to(logits.device, non_blocking=True)
if input_mask is not None:
input_mask = input_mask.to(logits.device, non_blocking=True)
masked_ids = input_ids.clone()
masked_ids[~input_mask] = -1
else:
masked_ids = input_ids
batch_sz, seq_len = masked_ids.shape
vocab_size = logits.size(-1)
if isinstance(p, Tensor):
penalty_per_row = p.expand(batch_sz, 1)
else:
penalty_per_row = torch.full(
(batch_sz, 1), float(p), device=logits.device, dtype=logits.dtype
)
counts = torch.zeros(
batch_sz, vocab_size, device=logits.device, dtype=logits.dtype
)
valid_mask = masked_ids >= 0
if valid_mask.any():
valid_ids = masked_ids[valid_mask]
row_indices = (
torch.arange(batch_sz, device=logits.device)
.unsqueeze(1)
.expand_as(masked_ids)[valid_mask]
)
counts.index_put_(
(row_indices, valid_ids),
torch.ones_like(valid_ids, dtype=logits.dtype),
accumulate=True,
)
return logits - penalty_per_row * counts
class SamplingPipeline(BaseSamplingStrategy):
"""Composes multiple sampling strategies into a single transformation.
Strategies are applied sequentially in the order they are provided,
matching the original temperature -> top-k -> top-p ordering.
Usage::
pipeline = SamplingPipeline([
TemperatureStrategy(0.8),
TopKStrategy(50),
TopPStrategy(0.95),
])
logits = pipeline.apply(logits)
token = pipeline.sample(logits) # softmax + multinomial
"""
def __init__(self, strategies: List[BaseSamplingStrategy]):
self.strategies = strategies
def apply(
self,
logits: Tensor,
filter_value: float = -float("inf"),
input_ids: Optional[Tensor] = None,
input_mask: Optional[Tensor] = None,
) -> Tensor:
for strategy in self.strategies:
logits = strategy.apply(logits, filter_value, input_ids, input_mask)
return logits
@staticmethod
def _is_greedy(temperature: Union[float, Tensor]) -> bool:
if isinstance(temperature, Tensor):
return bool((temperature == 0).all())
return temperature == 0
@torch.inference_mode()
def sample(
self,
logits: Tensor,
filter_value: float = -float("inf"),
input_ids: Optional[Tensor] = None,
input_mask: Optional[Tensor] = None,
return_logprobs: bool = False,
):
"""Apply strategies then sample (softmax + multinomial).
Short-circuits to ``argmax`` when temperature is exactly 0
(deterministic / greedy decode).
Args:
logits: Raw logits ``[batch, vocab_size]``.
input_ids: Previously generated token IDs ``[batch, seq_len]``.
input_mask: Boolean mask for ``input_ids`` padding.
return_logprobs: If ``True``, return ``(tokens, logprobs)``
where ``logprobs[i]`` is the log-probability of
``tokens[i]`` under the (post-strategy) sampling
distribution.
Returns:
Sampled token IDs ``[batch]``, or — when ``return_logprobs``
is ``True`` — a ``(token_ids, chosen_logprobs)`` tuple.
"""
if self._is_greedy_pipeline():
tokens = logits.argmax(dim=-1)
if not return_logprobs:
return tokens
log_probs = torch.log_softmax(logits.float(), dim=-1)
chosen = torch.gather(log_probs, -1, tokens.unsqueeze(-1)).squeeze(-1)
return tokens, chosen
transformed = self.apply(logits, filter_value, input_ids, input_mask)
tokens = torch.multinomial(
torch.softmax(transformed, dim=-1), num_samples=1
).squeeze(-1)
if not return_logprobs:
return tokens
log_probs = torch.log_softmax(transformed.float(), dim=-1)
chosen = torch.gather(log_probs, -1, tokens.unsqueeze(-1)).squeeze(-1)
return tokens, chosen
def _is_greedy_pipeline(self) -> bool:
"""True if the first strategy is greedy temperature (temp=0)."""
if not self.strategies:
return False
first = self.strategies[0]
return isinstance(first, TemperatureStrategy) and self._is_greedy(
first.temperature
)
@torch.inference_mode()
def sample(
logits: Tensor,
temperature: Union[float, Tensor] = 1.0,
top_k: Union[int, Tensor] = 0,
top_p: Union[float, Tensor] = 1.0,
frequency_penalty: Union[float, Tensor] = 0.0,
input_ids: Optional[Tensor] = None,
input_mask: Optional[Tensor] = None,
filter_value: float = -float("inf"),
return_logprobs: bool = False,
):
"""Apply sampling strategies then sample (softmax + multinomial).
Shortcut for ``SamplingPipeline(...).sample(logits, return_logprobs=)``.
When **temperature** is exactly 0 (scalar or single-element tensor)
the function short-circuits to ``argmax`` for deterministic decode.
When **frequency_penalty** is 0 (the common decode case), the entire
frequency penalty computation — including the O(batch * vocab) count
tensor allocation — is skipped.
Args:
logits: Raw logits ``[batch, vocab_size]``.
frequency_penalty: Penalty per occurrence for repeated tokens
(0.0 disables, range -2.0~2.0).
input_ids: Previously generated token IDs ``[batch, seq_len]``.
input_mask: Boolean mask for ``input_ids`` padding.
return_logprobs: If ``True``, also return the log-probability
of each sampled token under the (post-strategy) sampling
distribution — useful for RL rollout (PPO/GRPO importance
ratios).
Returns:
Sampled token IDs ``[batch]``, or — when ``return_logprobs`` is
``True`` — a ``(token_ids, chosen_logprobs)`` tuple where
``chosen_logprobs`` has shape ``[batch]``.
"""
has_freq = (
(isinstance(frequency_penalty, Tensor) and (frequency_penalty != 0).any())
if isinstance(frequency_penalty, Tensor)
else frequency_penalty != 0
)
strategies: List[BaseSamplingStrategy] = [
TemperatureStrategy(temperature),
TopKStrategy(top_k),
TopPStrategy(top_p),
]
if has_freq:
strategies.append(FrequencyPenaltyStrategy(frequency_penalty))
return SamplingPipeline(strategies).sample(
logits,
filter_value=filter_value,
input_ids=input_ids,
input_mask=input_mask,
return_logprobs=return_logprobs,
)