[diffusion] feat: add rollout `log_prob` with flow-matching SDE/CPS support by MikukuOvO · Pull Request #18806 · sgl-project/sglang

MikukuOvO · 2026-02-13T16:54:22Z

Motivation

This PR adds rollout log_prob support for diffusion flow-matching pipelines.
Previously, rollout paths did not expose consistent log_prob signals for flow-matching variants (especially SDE/CPS), which limited downstream training/evaluation workflows that depend on likelihood-based objectives.

Modifications

Add rollout log_prob computation in the diffusion rollout path.
Add flow-matching rollout support for SDE mode.
Add flow-matching rollout support for CPS mode.
Expose log_prob in rollout outputs with consistent shape/semantics across modes.
Keep backward compatibility for existing rollout callers when log_prob is not requested.
Add/extend related tests and validations for shape/dtype, mode coverage (SDE/CPS), and regression behavior.

Accuracy Tests

All tests were run with fixed random seeds.
Under the same seed and configuration, rollout outputs are deterministic: latent and log_prob are consistent across repeated runs.

Flux

Sampling modes tested: no rollout, CPS, SDE.
Verified deterministic consistency of both latent and log_prob under fixed seed.
Figure:

Qwen

Sampling modes tested: no rollout, CPS, SDE.
Verified deterministic consistency of both latent and log_prob under fixed seed.
Figure:

Z-Image

Sampling modes tested: no rollout, CPS, SDE.
Verified deterministic consistency of both latent and log_prob under fixed seed.
Figure:

Benchmarking and Profiling

No dedicated speed benchmark/profiling numbers are included in this PR yet.
Functional focus of this PR is rollout log_prob correctness and mode coverage.

Checklist

Format your code according to the Format code with pre-commit.
Add unit tests according to the Run and add unit tests.
Update documentation according to Write documentations.
Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
Follow the SGLang code style guidance.

Review Process

Ping Merge Oncalls to start the PR flow. See the PR Merge Process.
Get approvals from CODEOWNERS and other reviewers.
Trigger CI tests with comments or contact authorized users to do so.
After green CI and required approvals, ask Merge Oncalls to merge.

gemini-code-assist · 2026-02-13T16:54:47Z

Summary of Changes

Hello @MikukuOvO, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request significantly enhances the diffusion flow-matching pipeline by integrating the capability to compute and expose log_prob during the rollout process. This addition addresses a critical limitation where consistent log_prob signals were not available for flow-matching variants, particularly SDE and CPS modes. By providing these likelihood-based metrics, the change unlocks new possibilities for training and evaluation workflows that rely on such objectives, ensuring greater utility and flexibility for diffusion models.

Highlights

Rollout Log-Probability Support: Implemented log_prob computation during the diffusion rollout process, enabling likelihood-based objectives for downstream tasks.
Flow-Matching SDE/CPS Modes: Added specific support for Stochastic Differential Equation (SDE) and Conditional Probability Score (CPS) modes within the flow-matching rollout for log_prob calculation.
API and Parameter Exposure: Introduced new rollout and rollout_sde_type parameters in SamplingParams and exposed them through the OpenAI-compatible image and video generation APIs.
Trajectory Log-Probability Output: Ensured log_prob is consistently exposed in rollout outputs as trajectory_log_probs with appropriate shape and semantics across different modes.
Backward Compatibility: Maintained backward compatibility for existing rollout callers that do not explicitly request log_prob.

🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

Changelog

python/sglang/multimodal_gen/configs/sample/sampling_params.py
- Added rollout and rollout_sde_type fields to SamplingParams with default values.
- Extended add_cli_args to include command-line arguments for --rollout and --rollout-sde-type.
python/sglang/multimodal_gen/runtime/entrypoints/diffusion_generator.py
- Included trajectory_log_probs in the dictionary of output results for each prompt.
python/sglang/multimodal_gen/runtime/entrypoints/openai/image_api.py
- Added rollout and rollout_sde_type as optional parameters to _build_sampling_params_from_request.
- Passed rollout and rollout_sde_type to the SamplingParams construction.
- Included rollout and rollout_sde_type in the generations endpoint request.
python/sglang/multimodal_gen/runtime/entrypoints/openai/protocol.py
- Added optional rollout and rollout_sde_type fields to ImageGenerationsRequest.
- Added optional rollout and rollout_sde_type fields to VideoGenerationsRequest.
python/sglang/multimodal_gen/runtime/entrypoints/openai/video_api.py
- Conditionally added rollout and rollout_sde_type to sampling_kwargs in _build_sampling_params_from_request.
- Added rollout and rollout_sde_type as form parameters to the create_video endpoint.
- Passed rollout and rollout_sde_type to the VideoGenerationsRequest.
python/sglang/multimodal_gen/runtime/managers/gpu_worker.py
- Included trajectory_log_probs when constructing the OutputBatch from the result.
python/sglang/multimodal_gen/runtime/pipelines/patches/init.py
- Added an __init__.py file to the new patches directory.
python/sglang/multimodal_gen/runtime/pipelines/patches/flow_matching_with_logprob.py
- Added a new file containing _as_timestep_tensor and sde_step_with_logprob functions for computing log-probabilities in SDE and CPS modes.
python/sglang/multimodal_gen/runtime/pipelines_core/schedule_batch.py
- Added trajectory_log_probs as an optional torch.Tensor to the Req class.
- Added trajectory_log_probs as an optional torch.Tensor to the OutputBatch class.
python/sglang/multimodal_gen/runtime/pipelines_core/stages/decoding.py
- Passed trajectory_log_probs to the OutputBatch constructor.
python/sglang/multimodal_gen/runtime/pipelines_core/stages/decoding_av.py
- Passed trajectory_log_probs to the OutputBatch constructor.
python/sglang/multimodal_gen/runtime/pipelines_core/stages/denoising.py
- Imported sde_step_with_logprob from the new patches module.
- Patched sde_step_with_logprob onto the scheduler instance if not already present.
- Initialized an empty list trajectory_log_probs for collecting log probabilities.
- Added logic to determine rollout_enabled and rollout_sde_type, including warnings for invalid types.
- Conditionally used self.scheduler.sde_step_with_logprob to compute latents and step_log_prob when rollout is enabled, appending step_log_prob to the trajectory.
- Updated _post_denoising_loop to accept and process trajectory_log_probs, stacking them into a tensor if available.
python/sglang/multimodal_gen/runtime/pipelines_core/stages/denoising_dmd.py
- Initialized an empty trajectory_log_probs list when calling _post_denoising_loop.

Activity

The author, MikukuOvO, has implemented the core functionality for log_prob computation in diffusion flow-matching pipelines.
The author has integrated the new rollout and rollout_sde_type parameters across various configuration, API, and runtime files.
The author has added a new module flow_matching_with_logprob.py to handle the specific SDE/CPS log-probability calculations.
The author has performed accuracy tests across Flux, Qwen, and Z-Image models, verifying deterministic consistency of latent and log_prob under fixed seeds.
The author has noted that benchmarking and profiling are not yet included, focusing on functional correctness and mode coverage for this PR.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	`/gemini review`	Performs a code review for the current pull request in its current state.
Pull Request Summary	`/gemini summary`	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	`/gemini help`	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

gemini-code-assist

Code Review

This pull request introduces log_prob computation for rollouts in flow-matching diffusion pipelines, adding support for SDE and CPS modes. The changes are well-structured, correctly adding new sampling parameters, API endpoints, and plumbing the log_prob results through the pipeline stages. My review includes one suggestion to improve performance in the newly added sde_step_with_logprob function by optimizing a loop that could cause GPU-CPU synchronization overhead.

python/sglang/multimodal_gen/runtime/pipelines/patches/flow_matching_with_logprob.py

zhaochenyang20 · 2026-02-13T21:21:18Z

/rerun-failed-ci

zhaochenyang20 · 2026-02-14T20:23:47Z

/rerun-failed-ci

MikukuOvO · 2026-02-20T22:44:40Z

/rerun-failed-ci

zhaochenyang20 · 2026-02-21T05:16:20Z

/rerun-failed-ci

python/sglang/multimodal_gen/runtime/pipelines_core/stages/denoising.py

zhaochenyang20 · 2026-02-24T03:53:28Z

rebase and fix lint please.

python/sglang/multimodal_gen/runtime/pipelines/patches/flow_matching_with_logprob.py

zhaochenyang20 · 2026-02-24T21:31:07Z

Under the same seed and configuration, rollout outputs are deterministic: latent and log_prob are consistent across repeated runs.

This is nice. Do you think we can leverage latent and log_prob as metrics for CI?

zhaochenyang20

Fix the lint.
Could you add unit test to SDE and CPS. These APIs are important.

python/sglang/multimodal_gen/runtime/pipelines/patches/flow_matching_with_logprob.py

python/sglang/multimodal_gen/runtime/pipelines_core/stages/denoising.py

zhaochenyang20 · 2026-02-24T21:56:35Z

unit test could be like:

#19164 (comment)

…#19153 (sleep/wake) Cherry-picked from: - PR sgl-project#18806 (MikukuOvO): flow-matching SDE/CPS log_prob - PR sgl-project#19153 (Godmook): release/resume memory occupation Known issues: - t.item() in log_prob path causes GPU sync overhead - release_memory_occupation tags only supports "weights"

MikukuOvO · 2026-02-26T05:33:05Z

Under the same seed and configuration, rollout outputs are deterministic: latent and log_prob are consistent across repeated runs.

This is nice. Do you think we can leverage latent and log_prob as metrics for CI?

Great suggestion. Yes, I think we can leverage both latent and log_prob as CI metrics for rollout regression checks.

MikukuOvO · 2026-02-26T05:59:59Z

Thanks for the detailed review. I have addressed all the requested changes above (sync/device handling, comment/doc cleanup, and removal of dynamic getattr/hasattr usage in the touched paths).

I am currently working on adding unit tests for both SDE and CPS rollout paths, and will push the test updates next.

…#19153 (sleep/wake) Cherry-picked from: - PR sgl-project#18806 (MikukuOvO): flow-matching SDE/CPS log_prob - PR sgl-project#19153 (Godmook): release/resume memory occupation Known issues: - t.item() in log_prob path causes GPU sync overhead - release_memory_occupation tags only supports "weights"

zhaochenyang20 · 2026-02-27T03:51:17Z

/rerun-failed-ci

python/sglang/multimodal_gen/runtime/pipelines_core/schedule_batch.py

python/sglang/multimodal_gen/runtime/pipelines/patches/flow_matching_with_logprob.py

python/sglang/multimodal_gen/runtime/pipelines_core/stages/denoising.py

python/sglang/multimodal_gen/runtime/pipelines/patches/flow_matching_with_logprob.py

…t#18806)

…ob (sgl-project#18806)

MikukuOvO · 2026-03-02T05:01:30Z

Thanks for the reviews! I've gone through all your comments and pushed the fixes.

…port Rebased onto latest main.

zhaochenyang20 · 2026-03-03T02:31:43Z

This is my verification commands:

Install the changes:

cd python 
uv pip install -e ".[diffusion]"

With Python. Note that only python API can get log_probs.

from sglang import DiffGenerator

  gen = DiffGenerator.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers")

  # Mode 1: No Rollout (baseline)
  result_baseline = gen.generate(sampling_params_kwargs={
      "prompt": "A curious raccoon in a forest",
      "rollout": False, "save_output": True,
  })

  # Mode 2: SDE Rollout
  result_sde = gen.generate(sampling_params_kwargs={
      "prompt": "A curious raccoon in a forest",
      "rollout": True, "rollout_sde_type": "sde",
      "rollout_noise_level": 0.7, "return_trajectory_latents": True, "save_output": True,
  })

  # Mode 3: CPS Rollout
  result_cps = gen.generate(sampling_params_kwargs={
      "prompt": "A curious raccoon in a forest",
      "rollout": True, "rollout_sde_type": "cps",
      "rollout_noise_level": 0.7, "return_trajectory_latents": True, "save_output": True,
  })

  results = []
  for i in range(2):
      r = gen.generate(sampling_params_kwargs={
          "prompt": "A curious raccoon in a forest",
          "rollout": True, "rollout_sde_type": "sde",
          "rollout_noise_level": 0.7, "return_trajectory_latents": True, "seed": 42,
      })
      results.append(r)

  import torch
  print(f"Log_probs match: {torch.allclose(results[0].trajectory_log_probs, results[1].trajectory_log_probs)}")
  print(f"Latents match: {torch.allclose(results[0].trajectory_latents, results[1].trajectory_latents)}")

With Server

sglang serve --model-path Wan-AI/Wan2.1-T2V-1.3B-Diffusers --num-gpus 1

  # No Rollout                                                                                                                                         
  curl http://localhost:30000/v1/images/generations \             
    -H "Content-Type: application/json" \                                                                                                              
    -d '{                                                         
      "prompt": "A curious raccoon in a forest",
      "model": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
    }'

  # SDE Rollout
  curl http://localhost:30000/v1/images/generations \
    -H "Content-Type: application/json" \
    -d '{
      "prompt": "A curious raccoon in a forest",
      "model": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
      "rollout": true,
      "rollout_sde_type": "sde",
      "rollout_noise_level": 0.7
    }'

  # CPS Rollout
  curl http://localhost:30000/v1/images/generations \
    -H "Content-Type: application/json" \
    -d '{
      "prompt": "A curious raccoon in a forest",
      "model": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
      "rollout": true,
      "rollout_sde_type": "cps",
      "rollout_noise_level": 0.7
    }'

zhaochenyang20 · 2026-03-03T02:35:44Z

With Python. Note that only python API can get log_probs.

This is a critical issue. Every RL workload is running on a server, so the curl API should 100% have some ways to get log_probs and latent.

zhaochenyang20 · 2026-03-04T02:51:11Z

I do think that adding latent and log_prob in the response of the server is a problem.

LLM also has this requirement. Of course this is required.
The code - change is limited. I believe no more than 30 lines of code.
The latency effect is limited as well. If the user does not require these two fields, we do not send them back to the user. Then the latency is not affected.

Considering the latency of transfering large data via https, if you think that's slow, you are correct. But we still need to send them back. So,

Adds a request filed, like in your request, have a get_log_probs: True, get_latent: True. They are default to be False. But if true, transfer them back.
Evaluate the time for transferring these large data through the HTTPS server. The basic idea is copy what we do for LLM. Just check how LLM gets log_probs and so on. Do the same thing.

…port (sgl-project#18806) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

MikukuOvO · 2026-03-04T17:30:43Z

Quick manual verification (Qwen, Python API only)

cd python
uv pip install -e ".[diffusion]"

python - <<'PY'
import torch
from sglang.multimodal_gen import DiffGenerator

MODEL = "Qwen/Qwen-Image"
PROMPT = "A curious raccoon in a forest"

def one(gen, **kwargs):
    r = gen.generate(sampling_params_kwargs=kwargs)
    return r[0] if isinstance(r, list) else r

common = dict(
    prompt=PROMPT,
    save_output=False,              # quick test only
    rollout_noise_level=0.7,
    num_inference_steps=20,         # reduce runtime
)

with DiffGenerator.from_pretrained(model_path=MODEL, num_gpus=1) as gen:
    baseline = one(gen, **common, rollout=False)
    sde = one(gen, **common, rollout=True, rollout_sde_type="sde", return_trajectory_latents=True)
    cps = one(gen, **common, rollout=True, rollout_sde_type="cps", return_trajectory_latents=True)

    assert baseline.trajectory_log_probs is None
    assert sde.trajectory_log_probs is not None and sde.trajectory_latents is not None
    assert cps.trajectory_log_probs is not None and cps.trajectory_latents is not None

    r0 = one(gen, **common, rollout=True, rollout_sde_type="sde", return_trajectory_latents=True, seed=42)
    r1 = one(gen, **common, rollout=True, rollout_sde_type="sde", return_trajectory_latents=True, seed=42)

    lp_ok = torch.allclose(r0.trajectory_log_probs, r1.trajectory_log_probs)
    lat_ok = torch.allclose(r0.trajectory_latents, r1.trajectory_latents)
    assert lp_ok and lat_ok

    print("SDE log_probs shape:", tuple(sde.trajectory_log_probs.shape))
    print("SDE latents shape:", tuple(sde.trajectory_latents.shape))
    print("CPS log_probs shape:", tuple(cps.trajectory_log_probs.shape))
    print("CPS latents shape:", tuple(cps.trajectory_latents.shape))
    print("Determinism log_probs:", lp_ok)
    print("Determinism latents:", lat_ok)
    print("Quick rollout verification passed.")
PY

yhyang201 · 2026-03-06T03:31:12Z

@mickqian Nvidia CI passed and PR is approved, ready for merge

Updated docstring to clarify sde_type options.

zhaochenyang20 · 2026-03-06T23:32:37Z

python/sglang/multimodal_gen/runtime/pipelines/patches/flow_matching_with_logprob.py

+# SPDX-License-Identifier: Apache-2.0
+"""Flow-matching rollout step utilities for log-prob computation."""


If we adapt from other open-source diffusion workflows, we shall add acknowledgment here.

zhaochenyang20 · 2026-03-06T23:44:24Z

python/sglang/multimodal_gen/runtime/pipelines_core/stages/denoising.py

+                        if rollout_enabled:
+                            latents, step_log_prob = sde_step_with_logprob(
+                                self.scheduler,
+                                model_output=noise_pred,
+                                sample=latents,
+                                step_index=rollout_step_indices[i],
+                                generator=batch.generator,
+                                sde_type=rollout_sde_type,
+                                noise_level=rollout_noise_level,
+                            )
+                            trajectory_log_probs.append(step_log_prob)
+                        else:
+                            latents = self.scheduler.step(
+                                model_output=noise_pred,
+                                timestep=t_device,
+                                sample=latents,
+                                **extra_step_kwargs,
+                                return_dict=False,
+                            )[0]


It's a little bit unclear to me that sde_step_with_logprob vs self.scheduler.step in the input parameter. The most strange thing is that sde_step_with_logprob takes self.scheduler as a parameter while self.scheduler.step is an object method of scheduler. Could we share the same design pattern for parameters like:

change sde_step_with_logprob to self.scheduler.sde_step_with_logprob

Or, only have one entrypoint self.scheduler.step, but pass in step_index=rollout_step_indices[i], generator=batch.generator, sde_type=rollout_sde_type, noise_level=rollout_noise_level, as kwargs?

In deed I am not so sure about the process of SDE and CPS. Shall ask for help on design from BBuf, mick and Yuhao.

yhyang201 · 2026-03-07T00:49:13Z

@mickqian Nvidia CI passed and PR is approved, ready for merge

— SGLDHelper bot

alphabetc1 · 2026-03-09T18:43:39Z

python/sglang/multimodal_gen/runtime/pipelines_core/stages/denoising.py

-                            return_dict=False,
-                        )[0]
+                        if rollout_enabled:
+                            latents, step_log_prob = sde_step_with_logprob(


This rollout path bypasses self.scheduler.step(...) and directly computes the next sample from sigmas.

That seems not equivalent for multi-step schedulers like FlowUniPCMultistepScheduler, because their step() also updates internal state such as last_sample, model_outputs, timestep_list, and lower_order_nums.

Is this by design?

alphabetc1 · 2026-03-09T19:14:27Z

python/sglang/multimodal_gen/configs/sample/sampling_params.py

    save_output: bool = True
    return_frames: bool = False
+    rollout: bool = False
+    rollout_sde_type: str = "sde"


can we validate rollout_sde_type in _validate so the request fails early

Rockdu · 2026-03-23T08:49:08Z

Hi, since the original PR is relatively simple and lack necessary supports for parallel inference, I revamped the rollout part based on @MikukuOvO's version. See #21204
[Diffusion] Revamp Rollout Log-Prob Support with SDE/CPS for RL Post-Training

MikukuOvO requested review from mickqian and yhyang201 as code owners February 13, 2026 16:54

github-actions bot added the diffusion SGLang Diffusion label Feb 13, 2026

gemini-code-assist bot reviewed Feb 13, 2026

View reviewed changes

python/sglang/multimodal_gen/runtime/pipelines/patches/flow_matching_with_logprob.py Outdated Show resolved Hide resolved

zhaochenyang20 added the run-ci label Feb 13, 2026

MikukuOvO force-pushed the feat/rollout-logprob-support branch from 23595dd to 020befb Compare February 13, 2026 20:20

MikukuOvO requested a review from ping1jing2 as a code owner February 20, 2026 20:10

Rockdu reviewed Feb 21, 2026

View reviewed changes

python/sglang/multimodal_gen/runtime/pipelines_core/stages/denoising.py Outdated Show resolved Hide resolved

dreamyang-liu reviewed Feb 24, 2026

View reviewed changes

python/sglang/multimodal_gen/runtime/pipelines/patches/flow_matching_with_logprob.py Outdated Show resolved Hide resolved

dreamyang-liu reviewed Feb 24, 2026

View reviewed changes

python/sglang/multimodal_gen/runtime/pipelines/patches/flow_matching_with_logprob.py Outdated Show resolved Hide resolved

dreamyang-liu reviewed Feb 24, 2026

View reviewed changes

python/sglang/multimodal_gen/runtime/pipelines/patches/flow_matching_with_logprob.py Outdated Show resolved Hide resolved

zhaochenyang20 requested changes Feb 24, 2026

View reviewed changes

zhaochenyang20 mentioned this pull request Feb 24, 2026

[Feature] Establishing a Standardized Accuracy & Correctness Evaluation Suite for Diffusion #19164

Open

2 tasks

dreamyang-liu reviewed Feb 27, 2026

View reviewed changes

python/sglang/multimodal_gen/runtime/pipelines_core/schedule_batch.py Show resolved Hide resolved

dreamyang-liu reviewed Feb 27, 2026

View reviewed changes

python/sglang/multimodal_gen/runtime/pipelines/patches/flow_matching_with_logprob.py Outdated Show resolved Hide resolved

dreamyang-liu reviewed Feb 27, 2026

View reviewed changes

python/sglang/multimodal_gen/runtime/pipelines_core/stages/denoising.py Outdated Show resolved Hide resolved

dreamyang-liu reviewed Feb 27, 2026

View reviewed changes

python/sglang/multimodal_gen/runtime/pipelines/patches/flow_matching_with_logprob.py Outdated Show resolved Hide resolved

MikukuOvO added a commit to MikukuOvO/sglang that referenced this pull request Mar 1, 2026

[diffusion] fix: address rollout log_prob review feedback (sgl-projec…

dcaadb8

…t#18806)

MikukuOvO added a commit to MikukuOvO/sglang that referenced this pull request Mar 2, 2026

[diffusion] fix: precompute rollout step index and correct cps log_pr…

4abb27b

…ob (sgl-project#18806)

[diffusion] feat: add rollout log_prob with flow-matching SDE/CPS sup…

f1d30d1

…port Rebased onto latest main.

zhaochenyang20 force-pushed the feat/rollout-logprob-support branch from 4abb27b to f1d30d1 Compare March 3, 2026 02:07

Rockdu approved these changes Mar 4, 2026

View reviewed changes

This was referenced Mar 4, 2026

[Feature Request] Native SGLang-D Server API for Returning Extended Generation Metadata (Log Probs, Latents, etc.) #19827

Open

[Diffusion] Add PCG support for diffusion models #19828

Closed

Patch Up Diffusion Utils zhaochenyang20/sglang-diffusion-routing#36

Open

celve added a commit to celve/sglang that referenced this pull request Mar 4, 2026

[diffusion] feat: add rollout log_prob with flow-matching SDE/CPS sup…

5466192

…port (sgl-project#18806) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

MikukuOvO requested a review from zhaochenyang20 March 4, 2026 16:29

Godmook mentioned this pull request Mar 4, 2026

[Diffusion] Add native /v1/diffusion/generate endpoint for trajectory metadata #19892

Open

5 tasks

zhaochenyang20 mentioned this pull request Mar 5, 2026

[Bug] Memory not released after request log probs zhaochenyang20/sglang-diffusion-routing#38

Closed

Clarify sde_type options in docstring

19d822d

Updated docstring to clarify sde_type options.

zhaochenyang20 reviewed Mar 6, 2026

View reviewed changes

alphabetc1 reviewed Mar 9, 2026

View reviewed changes

Rockdu force-pushed the feat/rollout-logprob-support branch from 19d822d to 91e3f0d Compare March 23, 2026 07:54

Rockdu requested review from BBuf and yingluosanqian as code owners March 23, 2026 07:54

Rockdu force-pushed the feat/rollout-logprob-support branch 2 times, most recently from b0f98c8 to cd46eea Compare March 23, 2026 08:27

Rockdu force-pushed the feat/rollout-logprob-support branch from 3f0fa57 to 19d822d Compare March 23, 2026 09:41

		# SPDX-License-Identifier: Apache-2.0
		"""Flow-matching rollout step utilities for log-prob computation."""

Conversation

MikukuOvO commented Feb 13, 2026

Motivation

Modifications

Accuracy Tests

Flux

Qwen

Z-Image

Benchmarking and Profiling

Checklist

Review Process

Uh oh!

gemini-code-assist bot commented Feb 13, 2026

Summary of Changes

Highlights

Footnotes

Uh oh!

gemini-code-assist bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

Uh oh!

zhaochenyang20 commented Feb 13, 2026

Uh oh!

zhaochenyang20 commented Feb 14, 2026

Uh oh!

MikukuOvO commented Feb 20, 2026

Uh oh!

zhaochenyang20 commented Feb 21, 2026

Uh oh!

Uh oh!

zhaochenyang20 commented Feb 24, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

zhaochenyang20 commented Feb 24, 2026

Uh oh!

zhaochenyang20 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

zhaochenyang20 commented Feb 24, 2026

Uh oh!

MikukuOvO commented Feb 26, 2026

Uh oh!

MikukuOvO commented Feb 26, 2026

Uh oh!

zhaochenyang20 commented Feb 27, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

MikukuOvO commented Mar 2, 2026

Uh oh!

zhaochenyang20 commented Mar 3, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

zhaochenyang20 commented Mar 3, 2026

Uh oh!

zhaochenyang20 commented Mar 4, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

MikukuOvO commented Mar 4, 2026

Uh oh!

yhyang201 commented Mar 6, 2026

Uh oh!

zhaochenyang20 Mar 6, 2026

Choose a reason for hiding this comment

Uh oh!

zhaochenyang20 Mar 6, 2026

Choose a reason for hiding this comment

Uh oh!

yhyang201 commented Mar 7, 2026

zhaochenyang20 commented Mar 3, 2026 •

edited

Loading

zhaochenyang20 commented Mar 4, 2026 •

edited

Loading

alphabetc1 Mar 9, 2026 •

edited

Loading

Rockdu commented Mar 23, 2026 •

edited

Loading