[Fix] Account zero-logprob sequences correctly in chunked logprob stitching (#31639)

This commit is contained in:
Liangsheng Yin
2026-07-17 22:33:27 -07:00
committed by GitHub
parent ab3d421c30
commit 19c53c44a0
2 changed files with 171 additions and 30 deletions
+28 -30
View File
@@ -115,6 +115,12 @@ def get_token_ids_logprobs_raw(
vals.append([])
idxs.append([])
continue
if token_ids is None:
# The sequence's rows still occupy logprobs; step over them.
vals.append([])
idxs.append([])
pt += pruned_len
continue
token_ids_tensor = torch.tensor(token_ids, dtype=torch.long).to(
logprobs.device, non_blocking=True
)
@@ -169,10 +175,7 @@ def get_top_logprobs_chunk(
Returns:
int: Number of remaining tokens to process in next chunk
"""
# No sequences in the chunk
if logprobs.shape[0] == 0:
return 0
# Empty chunks still walk the slice to emit placeholder entries.
max_k = max(logits_metadata.top_logprobs_nums)
ret = logprobs.topk(max_k, dim=1)
values = ret.values.tolist()
@@ -208,13 +211,14 @@ def get_top_logprobs_chunk(
idx.append(indices[pt + j][:k])
# Append or extend based on whether the sequence was split across chunks
if len(val) > 0:
if split_pruned_len > 0:
input_top_logprobs_val[-1].extend(val)
input_top_logprobs_idx[-1].extend(idx)
else:
input_top_logprobs_val.append(val)
input_top_logprobs_idx.append(idx)
# Split-sequence continuations extend; everyone else owns a fresh
# (possibly empty) entry.
if split_pruned_len > 0:
input_top_logprobs_val[-1].extend(val)
input_top_logprobs_idx[-1].extend(idx)
else:
input_top_logprobs_val.append(val)
input_top_logprobs_idx.append(idx)
pt += pruned_len
return next_split_pruned_len
@@ -242,11 +246,7 @@ def get_token_ids_logprobs_chunk(
Returns:
int: Number of remaining tokens to process in next chunk
"""
# No sequences in the chunk
if logprobs.shape[0] == 0:
return 0
# Empty chunks still walk the slice to emit placeholder entries.
pt = 0
next_split_pruned_len = 0
for n, (token_ids, pruned_len) in enumerate(
@@ -280,14 +280,14 @@ def get_token_ids_logprobs_chunk(
val.append(logprobs[pt + j, token_ids].tolist())
idx.append(token_ids)
# Append or extend based on whether the sequence was split across chunks
if len(val) > 0:
if split_pruned_len > 0:
input_token_ids_logprobs_val[-1].extend(val)
input_token_ids_logprobs_idx[-1].extend(idx)
else:
input_token_ids_logprobs_val.append(val)
input_token_ids_logprobs_idx.append(idx)
# Split-sequence continuations extend; everyone else owns a fresh
# (possibly empty) entry.
if split_pruned_len > 0:
input_token_ids_logprobs_val[-1].extend(val)
input_token_ids_logprobs_idx[-1].extend(idx)
else:
input_token_ids_logprobs_val.append(val)
input_token_ids_logprobs_idx.append(idx)
pt += pruned_len
return next_split_pruned_len
@@ -541,19 +541,17 @@ class InputLogprobProcessor:
chunk_sample_indices = sample_indices[chunk_sample_mask] - start_idx
sampled_logits[chunk_sample_mask] = chunk_logits[chunk_sample_indices]
# If there are no input logprobs in this chunk, skip the rest
if chunk_indices.numel() == 0:
continue
# Zero-logprob-row chunks still need the per-sequence bookkeeping below.
# Compute the logprobs of the chunk
chunk_input_logprobs = chunk_logits[chunk_indices]
chunk_input_logprobs = torch.nn.functional.log_softmax(
chunk_input_logprobs, dim=-1
)
# For each chunk, we need to get the slice of the token_to_seq_idx
# End at the last row inside the chunk; token_to_seq_idx[end_idx]
# belongs to the next chunk and would emit its sequence twice.
chunk_slice = slice(
token_to_seq_idx[start_idx], token_to_seq_idx[end_idx] + 1
token_to_seq_idx[start_idx], token_to_seq_idx[end_idx - 1] + 1
)
# Get the logprob of top-k tokens
@@ -0,0 +1,143 @@
"""Chunked input-logprob processing must match the non-chunked reference.
Regression for the cross-chunk stitching accounting: zero-logprob-row
sequences (logprob opt-outs in mixed batches, mid-chunked-prefill segments)
were skipped or double-emitted, drifting the per-request entry counts that
the scheduler asserts on.
"""
import itertools
import unittest
from types import SimpleNamespace
import torch
from sglang.srt.layers.logprob_processor import InputLogprobProcessor
from sglang.test.ci.ci_register import register_cpu_ci
from sglang.test.test_utils import CustomTestCase
register_cpu_ci(est_time=30, suite="base-a-test-cpu")
VOCAB = 11
# Heterogeneous per-sequence parameters; uniform ones hide misalignment.
TOPK_CYCLE = [2, 0, 3]
# [] is a valid probe set distinct from None (opt-out).
TOKEN_IDS_CYCLE = [[0, 3], None, [1], []]
def _build_batch(seq_specs, with_token_ids):
"""seq_specs: list of (extend_len, logprob_start_len). Mirrors
LogitsProcessor._get_pruned_states for the extend-with-logprobs path."""
pruned_rows = []
token_to_seq_idx = []
sample_indices = []
input_logprob_indices = []
pruned_lens = []
sample_pt = -1
lp_pt = 0
for idx, (extend_len, start) in enumerate(seq_specs):
eff_start = start - 1 if extend_len == start else start
rows = extend_len - eff_start
pruned_rows.append(torch.randn(rows, VOCAB, dtype=torch.float32))
token_to_seq_idx.extend([idx] * rows)
sample_pt += rows
sample_indices.append(sample_pt)
n_lp = extend_len - start
input_logprob_indices.extend([lp_pt + i for i in range(n_lp)])
lp_pt += rows
pruned_lens.append(n_lp)
token_to_seq_idx.append(len(seq_specs) - 1)
metadata = SimpleNamespace(
extend_return_top_logprob=True,
extend_token_ids_logprob=with_token_ids,
top_logprobs_nums=[TOPK_CYCLE[i % 3] for i in range(len(seq_specs))],
extend_logprob_pruned_lens_cpu=pruned_lens,
extend_input_logprob_token_ids_gpu=torch.zeros(
len(input_logprob_indices), dtype=torch.int64
),
token_ids_logprobs=(
[TOKEN_IDS_CYCLE[i % len(TOKEN_IDS_CYCLE)] for i in range(len(seq_specs))]
if with_token_ids
else [None] * len(seq_specs)
),
)
return (
torch.cat(pruned_rows),
torch.tensor(sample_indices, dtype=torch.int64),
torch.tensor(input_logprob_indices, dtype=torch.int64),
token_to_seq_idx,
metadata,
)
def _run(proc, batch, chunked, chunk_size):
pruned_states, sample_indices, input_logprob_indices, t2s, metadata = batch
proc.enable_logprobs_chunk = chunked
proc.logprobs_chunk_size = chunk_size
def get_logits_fn(states, lm_head, logits_metadata, **kwargs):
return states.float()
return proc.forward(
pruned_states=pruned_states,
sample_indices=sample_indices,
input_logprob_indices=input_logprob_indices,
token_to_seq_idx=t2s,
lm_head=None,
get_logits_fn=get_logits_fn,
logits_metadata=metadata,
)
class TestLogprobChunkStitching(CustomTestCase):
def _sweep(self, with_token_ids):
torch.manual_seed(0)
proc = InputLogprobProcessor()
# (extend_len, start); start == extend_len is the degenerate
# zero-logprob-row shape.
menu = [(1, 1), (2, 2), (3, 0), (4, 1), (5, 5), (2, 0), (6, 2)]
tried = 0
for n_seqs in (1, 2, 3, 4):
for combo in itertools.product(menu, repeat=n_seqs):
batch = _build_batch(list(combo), with_token_ids)
# Same unit as the production gate: grid rows, not logprob rows.
total_rows = batch[0].shape[0]
for chunk_size in (1, 2, 3, 5):
if total_rows <= chunk_size:
continue
tried += 1
ref, ref_sampled = _run(proc, batch, False, 10**9)
got, got_sampled = _run(proc, batch, True, chunk_size)
label = f"specs={list(combo)} chunk={chunk_size}"
self.assertEqual(
ref.input_top_logprobs_val, got.input_top_logprobs_val, label
)
self.assertEqual(
ref.input_top_logprobs_idx, got.input_top_logprobs_idx, label
)
if with_token_ids:
self.assertEqual(
ref.input_token_ids_logprobs_val,
got.input_token_ids_logprobs_val,
label,
)
self.assertEqual(
ref.input_token_ids_logprobs_idx,
got.input_token_ids_logprobs_idx,
label,
)
torch.testing.assert_close(
ref.input_token_logprobs, got.input_token_logprobs, msg=label
)
torch.testing.assert_close(ref_sampled, got_sampled, msg=label)
self.assertGreater(tried, 1000)
def test_top_logprobs_stitching(self):
self._sweep(with_token_ids=False)
def test_token_ids_logprobs_stitching(self):
self._sweep(with_token_ids=True)
if __name__ == "__main__":
unittest.main()