Support DP sampling for spec decode by yubofredwang · Pull Request #232 · lightseekorg/tokenspeed

yubofredwang · 2026-05-24T01:42:19Z

Summary

Enable data parallele sampling. This has a few benefits:

No need to use deterministic kernel during larger batch. Potentially better perf.
Reduce the all gather communication overhead by calculating needed part and use all to all single operation.
Avoid duplicate work from ranks in TP mode. No need to sync verify results between ranks, but rather, gather the results.

Implemented:

swap vocab and batch dimension
onesided all2all kernel
padding for cuda graph and even split
Support logprobs
Support xgrammar

Test Plan

Ran Kimi-K2.5 TP8 FP8 KV, batch 40, EAGLE3 spec 5/1/6, ctx=5000, max_tokens=4096, comparing TP sampling baseline vs DP sampling onesided.

mode | round 1 tok/s | round 2 tok/s | round 3 tok/s | avg rounds 2-3
-- | -- | -- | -- | --
TP sampling | 4475.0 | 5227.5 | 4956.8 | 5092.2
DP sampling | 4333.1 | 5543.3 | 5238.8 | 5391.1

DP sampling was about +5.9% on warm rounds 2-3.

Run | Eval batch | DP gate | Route coverage | AIME score
-- | -- | -- | -- | --
Initial AIME | 16 | default min_bs=16 | bs=16 used DP sampling; smaller tail batches fell back | 0.9667
Large batch AIME | 30 | default min_bs=16 | bs=30 through bs=16 used DP sampling; bs<16 fell back | 0.9000
Full DP AIME | 30 | min_bs=1 | all decode batch sizes from bs=30 down to bs=1 used DP sampling | 0.9333

Follow up

Support better open threshold. Things that determine the minimum bsz to enable dp sampling are:

Batch size
Vocab size
TP size
spec draft number
DP attention enabled

chatgpt-codex-connector

💡 Codex Review

tokenspeed/python/tokenspeed/runtime/sampling/backends/flashinfer.py

Lines 522 to 524 in b940507

    
           write_output_logprobs( 
        
               logits_output, logits_output.next_token_logits, predict 
        
           )

Keep logits and token shapes aligned for output logprobs

In DP verify mode, predict is expanded to full-batch shape after gather_verify_outputs, but logits_output.next_token_logits is still this rank's local shard. If enable_output_logprobs is on, write_output_logprobs is called with mismatched row counts, which can fail or compute incorrect logprobs. The logprob write path needs full-batch logits (or a local-only token slice) before this call.

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 5b8440a7b7

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 16db036e43

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 9611a40417

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: d9c9d7ba91

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

Signed-off-by: Yubo Wang <yubowang2019@gmail.com> # Conflicts: # python/tokenspeed/runtime/execution/model_executor.py

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 779aacda7d

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

LorrinWWW · 2026-05-25T22:16:27Z

Do we need cuda_graph_wrapper.py and model_executor.py to be aware of DP sampling? I feel it may be cleaner to keep the DP sampling decision internal to logits_processor.py or the sampling wrappers, rather than exposing it to the higher-level execution path (I'm good to pass flags via controllers but feel like threshold decision etc can be internalized).

hmm I think we can't determine the graph bucket inside logits processor, because we need 1. forward path padding for other ops like MoE layer + 2. padding due to the bsz % tp_size

I'm confused:

What is "forward path padding for other ops like MoE layer"?

Could "padding due to the bsz % tp_size" be checked inside logits processor?

The blocker to moving this entirely into LogitsProcessor is that LogitsProcessor runs too late. By then we already needed to know:

DP sampling decision changes the logits/hidden-state communication contract. In current code, DP mode makes LogitsProcessor either slice hidden states before LM head when skip_all_gather is true, or use DP batch-vocab swap instead of normal TP logits all-gather. Non-DP uses the normal TP logits all-gather path.

But the deeper issue is that the decision must be made before logits processing so graph replay, logits layout, and sampling verify all agree.
@LorrinWWW please take a look at the new design to use a logitslayout to represent the decision

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: bb33395a62

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: b331b59ff5

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

…-sampling Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

yubofredwang · 2026-05-28T02:51:50Z

@syuoni I think the bigger issue is because LogitsProcessor needs to know whether to run all_gather or let sampling module run the all-to-all depending on dp sampling is enabled or not for this batch. Any suggestion?

Signed-off-by: Yubo Wang <yubowang2019@gmail.com> # Conflicts: # python/tokenspeed/runtime/sampling/backends/flashinfer.py

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 5ffca77a72

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

Signed-off-by: Yubo Wang <yubowang2019@gmail.com> # Conflicts: # python/tokenspeed/runtime/layers/logits_processor.py

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: cecc87207d

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-05-30T00:27:36Z

+    if forbid_global_logprob_writer:
+        import tokenspeed.runtime.sampling.backends.flashinfer as flashinfer_backend
+
+        original_writer = flashinfer_backend.write_output_logprobs


Remove the dead writer monkeypatch

In the enable_output_logprobs=True parametrization this branch is always taken, but tokenspeed.runtime.sampling.backends.flashinfer does not define or import write_output_logprobs anymore (checked repo-wide), so the test raises AttributeError here before it ever calls backend.verify. This means the new DP output-logprobs regression test cannot run in CI; either drop this monkeypatch or guard it against the current implementation.

Useful? React with 👍 / 👎.

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 6cfb53fb71

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-05-30T03:51:11Z

+        all_gather_into_tensor(
+            combined_full,
+            combined_local,
+            self._rank,


Pass the TP group, not rank, to all_gather

When the NCCL DP backend is used (either TOKENSPEED_DP_SAMPLING_BACKEND=nccl or auto on systems where the one-sided path is unavailable), this extra positional self._rank is bound as group and the real group is bound as backend, so the explicit backend= keyword raises TypeError: all_gather_into_tensor() got multiple values for argument 'backend' as soon as DP verify gathers outputs. The same call pattern is repeated for DP logprobs below; these calls should pass self._group as the third positional argument and omit the rank.

Useful? React with 👍 / 👎.

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: b9c62dbc68

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-05-30T04:00:37Z

+            dp_vocab_size = int(weight.shape[0]) * (
+                1 if processor.skip_all_gather else processor.tp_size
+            )


Pad replicated DP vocab before configuring comms

In the current patch the skip_all_gather branch now passes the replicated LM-head row count directly, but BaseCausalLM builds that head with config.vocab_size (not a TP-padded size). For TP+DP models whose vocab is not divisible by processor.tp_size, this value is then sent to configure_dp_sampling_vocab_size() and processor.configure_dp_sampling(), both of which construct DP comm helpers that require TP-divisible vocab sizes, so startup fails even though the replicated/full-vocab route does not need a vocab-shard swap. Please round/pad this value for the comm helpers or avoid constructing vocab-swap comm state when skip_all_gather is true.

Useful? React with 👍 / 👎.

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: d68b7313b3

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 4619fb6516

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: dd1dd0a8bc

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-06-01T04:57:16Z

+        dp_sampling_enabled=True,
+        dp_num_tokens_per_req=n,


Configure the DP logits executor before exercising DP logits

These GPU parity tests call processor._get_logits(..., dp_sampling=True) later, but passing dp_sampling_enabled=True to the constructor only flips the boolean and does not create _logits_layout_executor; the executor is created by configure_dp_sampling(). As written, the non-skip_all_gather DP path raises RuntimeError("dp_sampling logits layout executor is not configured") before the test can compare against legacy output, so the new regression tests cannot pass on a GPU runner.

Useful? React with 👍 / 👎.

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

yubofredwang requested a review from a team as a code owner May 24, 2026 01:42

chatgpt-codex-connector Bot reviewed May 24, 2026

View reviewed changes

Comment thread python/tokenspeed/runtime/sampling/backends/flashinfer.py Outdated

yubofredwang requested review from syuoni and yweng0828 May 24, 2026 06:55

chatgpt-codex-connector Bot reviewed May 24, 2026

View reviewed changes

Comment thread python/tokenspeed/runtime/sampling/backends/flashinfer.py

Comment thread python/tokenspeed/runtime/execution/drafter/eagle.py Outdated

yubofredwang force-pushed the ywang/dp-sampling branch 2 times, most recently from b940507 to 16db036 Compare May 24, 2026 08:29

chatgpt-codex-connector Bot reviewed May 24, 2026

View reviewed changes

Comment thread python/tokenspeed/runtime/execution/model_executor.py Outdated

chatgpt-codex-connector Bot reviewed May 24, 2026

View reviewed changes

Comment thread python/tokenspeed/runtime/sampling/backends/flashinfer.py Outdated

yubofredwang added 9 commits May 24, 2026 18:09

support dp sampling stage 1, cuda graph working

09e7289

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

onesided kernel

a2f92ae

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

new kernel, single cta barrier

5a10708

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

disable dp sampling by default

0a80023

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

code clean up and styling

a813284

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

fix test

7d557d2

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

Add DP sampling skip-all-gather regression coverage

9557840

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

Apply DP sampling lint formatting

857df62

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

test: clarify DP sampling metadata coverage

d9c9d7b

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

yubofredwang force-pushed the ywang/dp-sampling branch from 9611a40 to d9c9d7b Compare May 25, 2026 01:10

chatgpt-codex-connector Bot reviewed May 25, 2026

View reviewed changes

Comment thread tokenspeed-kernel/python/tokenspeed_kernel/ops/communication/dp_sampling.py Outdated

yubofredwang added 2 commits May 25, 2026 03:28

Merge remote-tracking branch 'origin/main' into ywang/dp-sampling

fc6b27e

Signed-off-by: Yubo Wang <yubowang2019@gmail.com> # Conflicts: # python/tokenspeed/runtime/execution/model_executor.py

Expose Kimi logits processor

779aacd

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

chatgpt-codex-connector Bot reviewed May 25, 2026

View reviewed changes

Comment thread python/tokenspeed/runtime/execution/model_executor.py Outdated

LorrinWWW reviewed May 25, 2026

View reviewed changes

Comment thread python/tokenspeed/runtime/execution/context.py Outdated

LorrinWWW reviewed May 25, 2026

View reviewed changes

yubofredwang added 2 commits May 26, 2026 04:15

Support DP verify masks and logprobs

4e6582c

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

Merge branch 'main' into ywang/dp-sampling

bb33395

chatgpt-codex-connector Bot reviewed May 26, 2026

View reviewed changes

Comment thread python/tokenspeed/runtime/distributed/dp_sampling_comm.py

Fix one-sided DP verify lazy init

b331b59

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

chatgpt-codex-connector Bot reviewed May 27, 2026

View reviewed changes

Comment thread python/tokenspeed/runtime/sampling/backends/flashinfer.py Outdated

Merge remote-tracking branch 'origin/ywang/dp-sampling' into ywang/dp…

d4310b3

…-sampling Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

syuoni reviewed May 28, 2026

View reviewed changes

Comment thread tokenspeed-kernel/python/tokenspeed_kernel/ops/communication/dp_sampling.py Outdated

yubofredwang added 3 commits May 29, 2026 01:08

Merge remote-tracking branch 'origin/main' into ywang/dp-sampling

a61dfa3

Signed-off-by: Yubo Wang <yubowang2019@gmail.com> # Conflicts: # python/tokenspeed/runtime/sampling/backends/flashinfer.py

Move DP sampling communication kernels under Triton ops

27cf114

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

Refactor DP sampling logits layout plan

5ffca77

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

chatgpt-codex-connector Bot reviewed May 29, 2026

View reviewed changes

Comment thread python/tokenspeed/runtime/execution/cuda_graph_wrapper.py Outdated

yubofredwang added 4 commits May 29, 2026 13:51

fix padding

7456ae6

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

Merge remote-tracking branch 'origin/main' into ywang/dp-sampling

2bfa5ce

Signed-off-by: Yubo Wang <yubowang2019@gmail.com> # Conflicts: # python/tokenspeed/runtime/layers/logits_processor.py

Rename DP sampling logits verify test

da79bd4

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

Test DP sampling graph bucket padding

cecc872

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

chatgpt-codex-connector Bot reviewed May 30, 2026

View reviewed changes

Merge branch 'main' into ywang/dp-sampling

6cfb53f

chatgpt-codex-connector Bot reviewed May 30, 2026

View reviewed changes

Fix DP verify logprobs regression test

b9c62db

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

chatgpt-codex-connector Bot reviewed May 30, 2026

View reviewed changes

Pad DP sampling comm vocab size

d68b731

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

chatgpt-codex-connector Bot reviewed May 30, 2026

View reviewed changes

Comment thread python/tokenspeed/runtime/layers/logits_processor.py

Remove redundant DP vocab sizing tests

4619fb6

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

chatgpt-codex-connector Bot reviewed May 30, 2026

View reviewed changes

Comment thread python/tokenspeed/runtime/execution/cuda_graph_wrapper.py Outdated

yubofredwang added 8 commits May 29, 2026 23:56

Reject DP sampling with sharded replicated-head path

2c40d13

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

Set DP sampling metadata for idle graph replay

d19427b

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

Polish DP logits layout plan setup

d49fc70

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

Merge branch 'main' into ywang/dp-sampling

5f8a1ba

Localize DP sampling layout policy

37a4f4c

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

Move DP sampling setup helpers into sampling

680b8c3

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

Move logits layout planner factory into sampling

9b6f449

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

Reduce execution coupling to DP sampling setup

dd1dd0a

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

chatgpt-codex-connector Bot reviewed Jun 1, 2026

View reviewed changes

Configure DP logits executor in parity test

d62a47b

Signed-off-by: Yubo Wang <yubowang2019@gmail.com>

	write_output_logprobs(
	logits_output, logits_output.next_token_logits, predict
	)

Conversation

yubofredwang commented May 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Test Plan

Follow up

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

Uh oh!

LorrinWWW May 25, 2026

Choose a reason for hiding this comment

Uh oh!

yubofredwang May 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

syuoni May 27, 2026

Choose a reason for hiding this comment

Uh oh!

yubofredwang Jun 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

Uh oh!

yubofredwang commented May 28, 2026

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot May 30, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

yubofredwang commented May 24, 2026 •

edited

Loading

yubofredwang May 26, 2026 •

edited

Loading

yubofredwang Jun 1, 2026 •

edited

Loading