[Trellis]Qwen3.5-35B; Weight Conversion and Raiden Weight Sync Integration by YixuanWang-99 · Pull Request #2083 · google/tunix

YixuanWang-99 · 2026-09-02T04:08:59Z

Description

Integrates Raiden weight synchronization with MaxText trainer and vLLM rollout workers for distributed RL workloads (GRPO / Trellis). Refactors destination-side weight sync into a shared mixin, handles multi-host FFI synchronization, manages cache lifecycles, and optimizes host memory usage.

Companion MaxText PR: AI-Hypercomputer/maxtext#5089

Key Changes

MaxText Trainer Integration (run_trainer_node.py, maxtext_utils.py): Adds trainer_backend="maxtext" support, configures mesh topologies (FSDP, TP, EP), enables weight conversion, and aligns trainer TP with rollout TP.
RaidenDestinationWeightSyncMixin Refactoring (raiden_weight_sync_mixin.py): Extracts a reusable mixin consolidating get_weight_sync_metadata, bind_weight_sync, sync_weights, and release_weight_sync across sampler adapters (inprocess_vllm_sampler_adapter, vanilla_sampler_adapter).
Multi-Host FFI & Step-0 Sync (raiden_handler.py, raiden_synchronizer.py, rl_program.py): Supports comma-separated multi-host control_plane_rpc_address registrations, forces weight sync at step 0 to prevent uninitialized generation, and adds tensor/element count verification to checksum diagnostics.
Cache Lifecycle & Memory Optimization (raiden_weight_sync_delegate.py): Adds prefix cache reset and KV cache reinitialization; adds --enable_prefix_caching flag (default false) to prevent recurrent state desync in hybrid models; releases host arrays post-sync.
128-Lane MoE Chunking & Launch Script (tunix.generate.utils, launch_raiden.sh): Aligns MoE interleaving to 128-element lane chunks for TPU GMM v2 kernels, and provides unified cluster deployment and triage automation.

Testing

Unit Tests: Updated and verified raiden_handler_test.py, raiden_synchronizer_test.py, and weight_sync_test.py.
E2E Verification: Validated full 2-step distributed RL loop on Cloud TPU v5p with MaxText trainer and vLLM rollout workers, logs

Checklist

Unit tests added and verified.
Verified compatibility with companion MaxText PR.
Verified on multi-node TPU cluster.

The distributed runtime imports discovery_service_pb2 and discovery_service_pb2_grpc, but only the .proto was tracked. That works in an editable install, where the generated files sit in the working tree, and fails anywhere the package is installed from a GitHub archive -- an archive contains only tracked files, so the stubs are absent and every worker process dies at import: ImportError: cannot import name 'discovery_service_pb2' from 'tunix.experimental.distributed.runtime.discovery' (unknown location) Observed on all three worker pods of a GKE run. Nothing in the packaging generates them: there is no protoc step in pyproject.toml, and the .proto is not shipped in the wheel either, so generating at install time is not an option without also packaging the source. Generated with grpcio-tools 1.81.1 / protobuf 6.33.6, which stamps gencode 6.33.5. The gencode version must not exceed the protobuf runtime wherever these execute, and the post-training image ships protobuf 6.33.6; stubs built with a 7.x toolchain load fine locally but abort there with VersionError: Detected incompatible Protobuf Gencode/Runtime versions ... gencode 7.35.1 runtime 6.33.6 Regenerate with a protobuf 6.x toolchain, from the repo root so the imports are fully qualified.

CPU_MACHINE, TPU_SLICE and GCS_SCRATCH_LOCATION were hardcoded to what trellis-demo-0810 happened to have, so the launcher only ran on that cluster. mlperf-v5p has no n2-standard-64 pool and no 2x2x2 topology, and Pathways' default compilation-cache bucket is not writable from this project -- which kills the compilation service rather than degrading it. All three are now ${VAR:-<previous default>}, so existing invocations are unchanged. The jobsets gain HF_TOKEN from a secretKeyRef. tunix/oss/utils.py calls hf.login() whenever HF_TOKEN is unset, even for a public model, which in a pod polls for a device code until it fails. Marked optional: true so pods without the secret still start. The secret name is still hardcoded to anisha-hf-token and should be parameterised before this goes anywhere beyond the current cluster.

maxtext_engine.py's release_weight_sync() docstring already claimed to release staged weight buffers after transfer completion, but the implementation only logged metrics -- self.arrays (the host-staged copy) kept lingering for the entire idle window between rounds with nothing to actually free it. This alone does not shrink the native transport's own hold lifetime (BindWeights only releases the previous round's hold atomically with acquiring the next one -- see the chunking fix in maxtext, commit 7f90d9c53, for the actual peak-memory fix), but it keeps our own Python-side reference from outliving its purpose once d2h() has already copied the data into the native transport's persistent buffer, and it makes the intent of release_weight_sync() match what its docstring already claimed. self.names is left untouched so `bound` keeps reporting whether bind() has ever run. (cherry picked from commit 1fc6ff9)

Ports vllm_sampler_adapter.py from mohit/raiden-maxtext-rlvllm, which drives tpu-inference's RLVllmSampler (SAMPLER=vllm) instead of running a vLLM engine in-process. Imports are repointed at the weight_sync package, since the orchestrator module this branch was written against has since been split out. That repointing is also why dict_to_metadata comes along here: it lives in orchestrator/weight_sync.py on the source branch and had no counterpart under weight_sync/, so the adapter resolved its module but not the symbol, and only failed once weight sync round 0 actually called it. The two paths need it for different reasons -- an in-process destination returns WorkUnitMetadata directly, but this one binds Raiden inside a separate EngineCore process, so tpu-inference flattens the same content to plain dicts to cross that boundary. Rebuilding the dataclasses here lets both present one type to manifest preflight.

Four fixes, each needed before the rollout node would start under --sampler=vllm. All were found by running it; none are speculative. Import tpu_sync's native extension first. Loading it after vLLM/torch pull in their own copy aborts the process with "free(): invalid pointer" inside tpu_inference.rl.raiden_worker_sync, which imports the same module. Bisecting the three sampler modes showed the crash is independent of weight sync, so the guarded import goes above every other import and is load-bearing rather than stylistic. Resolve --tensor_parallel_size after parsing instead of defaulting it to jax.device_count(). Evaluating that default opens the TPU in the parent while the parser is still being built; the vllm path then hands those same chips to a separate EngineCore process, which either cannot reopen them ("Device or resource busy") or inherits them half-owned via fork and hangs on its first compile. The launcher already says which chips we own, so read that and only fall back to JAX when it is unset. Set NEW_MODEL_DESIGN for the MaxText branch. MaxText's inference vllm.yml declares a five-axis mesh, but tpu-inference only builds one under that env; its default 2D ('data','model') mesh fails model loading with "Resource axis: attn_dp ... is not found in mesh". Scoped to this branch so the HF path keeps the 2D mesh its MoE kernel wants. Thread MAXTEXT_MODEL_NAME from one launcher variable into both the trainer and the rollout. Defaulting the two sides separately is how they drift, and drift here is not a clean failure: Raiden pairs tensors by exact name, so a MaxText trainer against a non-MaxText rollout matches zero names.

The failures ride on WeightSyncError structurally, but the default renderer prints only the outer message, so a mismatch reported "preflight failed" without naming anything. Problems are name-sorted, so a whole-convention mismatch fills the head of the list with one side only; sample both manifests explicitly, since that is what identifies the convention gap and the problem lines alone cannot show it.

Nothing on this path records the completion, so a run that generates nothing but newlines reports the same "rollouts=N" summary as a real one -- which is exactly how a broken Mode 2 rollout was mistaken for a working one. Logs token counts and a text prefix per trajectory. Revert once the vllm sampler path is trusted; this is a debugging aid, not a feature.

train_stage only syncs after a step, so step-0 trajectories came from whatever the rollout worker built for itself. On the MaxText-in-vLLM path that is a randomly initialized model: MaxTextForCausalLM is constructed with an empty load_parameters_path, so from_pretrained skips the Orbax restore. The resulting degenerate rollouts were still scored and trained on, so the run looked healthy while learning from noise. StandardRLProgram now pushes the trainer's starting weights out before dispatching anything, retrying until the vLLM EngineCore subprocess comes up (~20s locally) and failing loudly at the deadline rather than rolling out from uninitialized weights. That sync needs the engine, which RLVllmSampler only builds lazily on the first sample(), so VllmSamplerAdapter starts it on demand from the two weight-sync entry points. Without that the round finds no worker, reports an empty destination manifest, and the run deadlocks with the engine waiting for a sample that dispatch is waiting on the sync to allow.

The source and destination grand totals are only comparable if both sides bound the same tensors, which the checksum dict left implicit. Emitting __tensor_count__ and __element_count__ alongside makes that checkable rather than assumed; tpu-inference's RaidenWorkerSync.checksums() reports the same keys.

…istributed configs - In RaidenSynchronizer, flatten state with incremental device-to-cpu transfers, GC, and malloc_trim to minimize peak host memory - Support flat_state attribute in transfer_state_directly for nnx state updates - Make is_bounded check idempotent in sampler adapters - Increase PhaseTimeouts bind and metadata limits to 180s - Add CLI options and auto MoE padding calculation for MaxText trainer/rollout nodes

…ode.py - Set vllm.use_weight_converter and vllm.rollout_backend in MaxText config argv in run_trainer_node.py

…depool toleration - Add gke-nodepool tolerations in jobset.cpu.yaml - Replace TP arguments with --tensor_parallel_size in run_rollout_node.py - Use thread pool executor for submit_worker in DistributedRLEngine when loop is running - Strip .value suffix from variable names in VllmSamplerAdapter, flatten_weights, and WorkUnitMetadata - Support host_stage mode in RaidenSynchronizer under proxy backend

…th Raiden sync - Provide one-command launcher, monitor, triage, and teardown for MaxText + Tunix + vLLM + Raiden workloads - Support model presets (qwen3-0.6b, qwen3.5-35b, qwen3-1.7b) with verified slice topologies and checkpoints - Include automated log triaging for weight conversion and hardware errors

…ch_raiden.sh

…mands in launch_raiden.sh - Truncate default RUN_ID username prefix to stay within Kubernetes 63-character label value limit for coordinator labels - Add explicit length check rejecting TRAINER_ID > 26 characters - Add start-trainer and start-rollout subcommands - Handle render and dry-run commands cleanly - Wait for JobSet deletion during workload teardown

…replica unit ID - Thread --rollout_replicas from launch_raiden.sh to run_gsm8k_dist_grpo.py - Wait for all rollout worker replicas in cluster.wait_for_workers() - Set unit=server_id in VllmSamplerAdapter.get_raiden_metadata() to disambiguate replicas - Eagerly drain completed responses in DistributedRLEngine.poll_rollouts() with non-blocking poll loop

…date test assertions - Dynamically register custom call targets via libraiden_ffi_bridge.so in RaidenSynchronizer - Parameterize HF_TOKEN_SECRET_NAME in yaml_generator and jobset templates with maxRestarts=3 - Deduplicate dict_to_metadata in weight_sync.py - Update checksums test assertions in raiden_synchronizer_test.py for tensor and element counts - Clean up orphaned batch jobs and default SYNC_CODE=true in launch_raiden.sh

…oyment configuration - Support comma-separated multi-axis sharding specifications and monkey-patch TPUWorker in raiden_synchronizer - Parameterize k8s namespace and improve status/log triage in launch_raiden.sh and deployment YAMLs - Respect DISABLE_CHECKPOINTING environment variable in run_trainer_node and rl_program - Add base_num_kv_heads configuration parameter in maxtext_utils - Lower bind and metadata timeouts and improve shard metadata logging in weight_sync_coordinator

…nable MoE weight prefusion, and support multi-host FFI synchronization

… rollout and trainer

…efault: false) to prevent recurrent state desync in hybrid models

…nHandler registration

…lifecycle management - Extract RaidenDestinationWeightSyncMixin for rollout sampler adapters - Support prefix cache reset and KV cache deletion/reinitialization in RaidenWeightSyncDelegate - Release host arrays after weight sync in PeftTrainer - Guard patch_raiden_worker_sync when JAX_PLATFORMS=cpu - Define TPU_V5P_SUBCORE_LANE_SIZE constant in tunix.generate.utils - Remove stale launch_raiden.sh symlink

…ndant launch_raiden.sh

s-noghabi · 2026-09-08T20:11:49Z

why do we want to add these? these should be auto generated on setup.

+1 please revert

s-noghabi · 2026-09-08T20:13:34Z

    completion = action.action if hasattr(action, "action") else str(action)
    reward, info = gsm8k_env_reward(self.task, action)
    info["correct"] = bool(info["answer_correct"])
+    prompt_text = self.task.get("question", "") or self.task.get("prompts", "")


this is debugging logs, we should push all these. especially at info level. For this and other changes, can we do a thorough review to make sure we don't clutter the logs

s-noghabi · 2026-09-08T20:15:30Z

      return self._trainer.prepare_weight_sync(**kwargs)

  def save_checkpoint(self, metadata: Any = None, **kwargs) -> None:
+    if os.environ.get("DISABLE_CHECKPOINTING", "false").lower() in ("true", "1"):


instead of os env checking can we pass through the config?

s-noghabi · 2026-09-08T20:15:57Z


  async def poll_rollouts(
-      self, timeout_s: float = remote_execution.LONG_POLL_TIMEOUT_S
+      self, timeout_s: float = 1.0


why are we changing these values? Also in a few other locations.

timeout of 1s doesn't make any sense, please revert

s-noghabi · 2026-09-08T20:20:50Z

+      if res is None or isinstance(res, Exception):
+        return []
+      worker_items = [res]
+      while True:


why do we need this second while loop? the polling automatically checks in a loop too.

s-noghabi · 2026-09-08T20:21:16Z

-                    "num_microbatches": num_microbatches,
-                },
-            )
+            if os.environ.get("DISABLE_CHECKPOINTING", "false").lower() in ("true", "1"):


ditto, let's pass this properly

s-noghabi · 2026-09-08T20:21:27Z


  async def poll_rollouts(
-      self, timeout_s: float = remote_execution.LONG_POLL_TIMEOUT_S
+      self, timeout_s: float = 1.0


s-noghabi · 2026-09-08T20:23:10Z

+  async def _sync_initial_weights(self) -> None:
+    """Pushes the trainer's starting weights out before the first dispatch.
+
+    `train_stage` only syncs *after* a step, so whatever the rollout worker


shouldn't this change be pushed to MaxText-in-vLLM instead so uit doesn't load from an empty checkpoint? that seems like a better level of abstraction and modularity.

s-noghabi · 2026-09-08T20:24:16Z

      help="Weight sync mode (none, fallback, or raiden).",
  )
-  return parser.parse_args(argv)
+  parser.add_argument(


can these improvement be split into their own cl? doesn't seem relevant for the fix iiuc.

s-noghabi · 2026-09-08T20:24:38Z

          " step 0)."
      ),
  )
+  parser.add_argument(


ditto on seperate cl

tianshub

Please split this PR in the smaller ones and make sure each one contains a well focused change

tianshub · 2026-09-08T20:41:53Z

+1 please revert

tianshub · 2026-09-08T20:42:33Z

 #!/bin/bash
 # Copyright 2026 Google LLC
 #
-# Licensed under the Apache License, Version 2.0 (the "License");


this banner should still be there

tianshub · 2026-09-08T20:43:30Z

@@ -1,107 +1,889 @@
 #!/bin/bash


why do we need all these changes in the launcher script? it was runnable previously no?

tianshub · 2026-09-08T20:45:04Z


  async def poll_rollouts(
-      self, timeout_s: float = remote_execution.LONG_POLL_TIMEOUT_S
+      self, timeout_s: float = 1.0


timeout of 1s doesn't make any sense, please revert

tianshub · 2026-09-08T20:46:30Z

  )


+def _submit_worker(


is any of the changes in this file related to weight sync? also the asyncio loop should already be running when we start invoke worker endpoints, why do we need to handle this here?

tianshub · 2026-09-08T20:47:46Z

        self.on_step_end(current_step, step_result)
      self._step += 1

+  async def _sync_initial_weights(self) -> None:


I'm pretty sure I've seen this function somewhere in another PR, do you need to rebase and merge?

tianshub · 2026-09-08T20:48:41Z

    self.max_response_length = request.generation_kwargs.get(
        "max_response_length"
    )
+    self._accumulated_token_ids: List[int] = []


why do we need these changes?

tianshub · 2026-09-08T20:50:01Z

          " transfer_parallelism, not both"
      )
-    worker_rpc_client = None
-    if name_resolver is not None:


why is this if condition removed? weight sync is not a mandatory arg for trainer

- Point PRESET_DEFAULT_IMAGE to yixuann-dev-0908-ffi for qwen35-35b-moe-raiden - Add flag aliases for --model-name/--model_name and --queue-name/--queue_name

YixuanWang-99 · 2026-09-09T01:23:26Z

Thanks everyone for the detailed review feedback!

Per the feedback on PR size and scope, this PR is being split into a stacked chain of focused, reviewable PRs targeting main:

Stacked PR Chain:

[Raiden Weight Sync 1/6] Support multi-host discovery and multi-axis sharding in RaidenHandler #2146 (T1): Multi-host discovery & multi-axis sharding in RaidenHandler (raiden_handler.py + tests)
- Restores the name_resolver is not None guard (@tianshub).
[Raiden Weight Sync 2/6] FFI transport selection, preflight mismatch check, and host-stage normalization #2147 (T2a): FFI transport selection, preflight mismatch check, and host-stage normalization (raiden_synchronizer.py, weight_sync_coordinator.py, weight_sync.py)
- Fixes the Step 0 blocker where FFI was never enabled end-to-end between Pathways trainer and rollout. Adds preflight validation in WeightSyncCoordinator to reject mismatched transport modes before transfer.
- Pairs with MaxText #5171 (M4).
[Raiden Weight Sync 3/6] Extract RaidenDestinationWeightSyncMixin for rollout samplers #2148 (T3): Extract RaidenDestinationWeightSyncMixin for rollout samplers
- Clean net −65 line refactoring extracting common destination lifecycle methods across inprocess_vllm, vanilla, and external vLLM adapters.
[Raiden Weight Sync 4/6] Manage cache lifecycle and prefix caching in RaidenWeightSyncDelegate #2149 (T4): Cache lifecycle management and prefix caching
- Prefix cache invalidation + KV state re-init upon weight transfer. Adds --enable_prefix_caching flag (default false) to avoid recurrent state desync in hybrid models.
[Raiden Weight Sync 5/6] MoE 128-lane weight interleaving for TPU GMM layout #2150 (T5): MoE 128-lane weight interleaving for TPU GMM layout (tunix/generate/utils.py)
- Matches TPU GMM kernel interleaving (TPU_V5P_SUBCORE_LANE_SIZE = 128).
- Pairs with MaxText #5168 (M2). Includes dedicated unit tests.
[Raiden Weight Sync 6/6] MaxText trainer config plumbing and MoE dimension padding #2151 (T6): MaxText trainer configuration plumbing (maxtext_utils.py)
- Replaces ad-hoc env-var checking with explicit parameters (base_num_kv_heads, rollout_mesh_tp, prefuse_moe_weights, padded_moe_mlp_dim).
[Deployment] Parameterize JobSet manifests and minimal k8s_launcher for multi-host RL #2152 (T9): Parameterize JobSet manifests and update k8s_launcher.sh
- Independent deployment slice. Restores Apache 2.0 license banner (@tianshub).

Notes on review comments & dropped duplicates:

Duplicate discovery pb2 stubs: Dropped in favor of [Tunix] Support Raiden weight sync on CPU-staged Pathways arrays and add distributed Qwen3-32B launcher #2119 (@andytwigg).
1s timeouts & inner poll loop in distributed_rl_engine.py: Reverted to standard timeouts; dropped _submit_worker.
Debugging log dumps in gsm8k.py: Reverted.
**T7 (Step-0 weight sync in rl_program.py)**: Root cause confirmed in MaxText rollout model initialization (MaxTextForCausalLM skipping restore on empty path). T7 is held from this chain and preserved on yixuann-debug-raiden.

We will keep yixuann-debug-raiden alive as the integration branch while landing the stacked PRs.

YixuanWang-99 requested review from abheesht17, hgao327, jiangyangmu, lc5211, s-noghabi, sizhit2, tianshub and wang2yn84 as code owners September 2, 2026 04:09

github-actions Bot assigned lc5211 Sep 2, 2026

YixuanWang-99 force-pushed the yixuann-debug-raiden branch 2 times, most recently from 3a7fe34 to 6e08b14 Compare September 3, 2026 19:13

A9isha and others added 11 commits September 3, 2026 19:18

Address review feedback: explicitly pass vllm config in run_trainer_n…

c46b832

…ode.py - Set vllm.use_weight_converter and vllm.rollout_backend in MaxText config argv in run_trainer_node.py

YixuanWang-99 force-pushed the yixuann-debug-raiden branch from 6e08b14 to c46b832 Compare September 3, 2026 20:13

YixuanWang-99 added 3 commits September 3, 2026 22:17

Update default container image to yixuann-debug-raiden-0903-2 in laun…

a831deb

…ch_raiden.sh

YixuanWang-99 changed the title ~~[DO NOT REVIEW][Trellis]Weight conversion that compatible with raiden~~ [Trellis]Weight conversion that compatible with raiden Sep 3, 2026

YixuanWang-99 added 3 commits September 3, 2026 23:39

Update default container image to yixuann-raiden-debug-0903-2 in laun…

5af6e07

…ch_raiden.sh

YixuanWang-99 added 8 commits September 4, 2026 07:25

Fix Qwen3.5-35B Raiden weight sync: align trainer TP to rollout TP, e…

8ff76ca

…nable MoE weight prefusion, and support multi-host FFI synchronization

Add --use-weight-converter CLI flag support in launch_raiden.sh

bee926f

Improve code quality and error handling for Raiden weight sync across…

d7f98e9

… rollout and trainer

Fix MoE 128-lane interleaving and add --enable_prefix_caching flag (d…

1888dd8

…efault: false) to prevent recurrent state desync in hybrid models

Support multi-host comma-separated control_plane_rpc_address in Raide…

2ad3bd8

…nHandler registration

YixuanWang-99 changed the title ~~[Trellis]Weight conversion that compatible with raiden~~ [Trellis]Weight Conversion and Raiden Weight Sync Integration Sep 8, 2026

YixuanWang-99 changed the title ~~[Trellis]Weight Conversion and Raiden Weight Sync Integration~~ [Trellis]Qwen3.5-35B; Weight Conversion and Raiden Weight Sync Integration Sep 8, 2026

YixuanWang-99 mentioned this pull request Sep 8, 2026

[Trellis]Qwen3.5-35B; Weight Conversion and Raiden Weight Sync Integration AI-Hypercomputer/maxtext#5089

Open

4 tasks

YixuanWang-99 added 2 commits September 8, 2026 19:05

Log sampled rollout prompt and response texts in GSM8KEnv._step_impl

3966b80

Consolidate e2e launcher scripts into k8s_launcher.sh and remove redu…

72bbc1c

…ndant launch_raiden.sh

s-noghabi requested changes Sep 8, 2026

View reviewed changes

Update default runner image to yixuann-dev-0908

6c36550

tianshub requested changes Sep 8, 2026

View reviewed changes

Update default runner image and add flag aliases in k8s_launcher.sh

8807d63

- Point PRESET_DEFAULT_IMAGE to yixuann-dev-0908-ffi for qwen35-35b-moe-raiden - Add flag aliases for --model-name/--model_name and --queue-name/--queue_name

Uh oh!

Conversation

YixuanWang-99 commented Sep 2, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Key Changes

Testing

Checklist

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

tianshub left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

YixuanWang-99 commented Sep 9, 2026

Stacked PR Chain:

Notes on review comments & dropped duplicates:

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

YixuanWang-99 commented Sep 2, 2026 •

edited

Loading