Skip to content

Fix sample patch #333

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
jinminxi104 merged 4 commits into from
Jun 27, 2026
Merged

Fix sample patch #333

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
f174a6c
fix vision in qwen35
jinminxi104 Jun 9, 2026
672f4cf
fix sampling patch
jinminxi104 Jun 24, 2026
ff13711
Merge branch 'main' into fix_sample_patch
jinminxi104 Jun 25, 2026
c5f3a3b
lint
jinminxi104 Jun 25, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
32 changes: 29 additions & 3 deletions dlinfer/framework/lmdeploy_ext/device/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,6 +2,8 @@
import importlib
import torch
from functools import lru_cache
from lmdeploy.pytorch.model_inputs import ModelInputs
from lmdeploy.pytorch.strategies.base.model_agent import ExtraInputs
from dlinfer.vendor import vendor_name

vendor = ["camb", "ascend"]
Expand Down Expand Up @@ -73,9 +75,24 @@ def patch_async_sampling_logits():
)

async def async_sampling_logits(
self, logits: torch.Tensor, sampling_inputs: SamplingInputs
self,
logits: torch.Tensor,
inputs: ModelInputs,
extra_inputs: ExtraInputs,
sampling_inputs: SamplingInputs,
):
"""Sampling logits."""
if self.spec_agent.is_enabled():
extra_inputs.target_logits = extra_inputs.target_logits.to(torch.float32)
extra_inputs = await self.spec_agent.async_sampling_logits(
inputs, extra_inputs, sampling_inputs
)
return (
extra_inputs.next_token_ids,
extra_inputs.logprobs,
extra_inputs.output_token_ids,
extra_inputs,
)
# record function does not support async function
# so we can not decorate it on async_sampling_logits
with record_function("sampling_logits"):
Expand All @@ -88,14 +105,18 @@ async def async_sampling_logits(
origin_logits = logits
logits, raw_logprobs = await logits_processor(origin_logits)
next_token_ids = logits_processor.sampling(logits)
await logits_processor.accept_guided_tokens(next_token_ids)
logprobs = logits_processor.compute_logprobs(raw_logprobs, next_token_ids)
if logprobs is not None:
logprobs = BatchedLogProbs(
vals=logprobs[0],
indices=logprobs[1],
)

return next_token_ids, logprobs
# post sampling
next_token_ids, extra_inputs = self.agent_strategy.post_sampling(
inputs, logits, next_token_ids, extra_inputs
)
return next_token_ids, logprobs, next_token_ids, extra_inputs

BaseModelAgent.async_sampling_logits = async_sampling_logits

Expand Down Expand Up @@ -608,6 +629,11 @@ def custom_prepare_inputs_for_generation(
pixel_values is not None or context.is_chunk_multimodal
)

# return input embeds for spec decoding
return_input_embeds = self.is_spec_decoding and (
pixel_values is not None or context.is_chunk_multimodal
)

# inputs of forward
return dict(
input_ids=input_ids,
Expand Down
Loading