Correctness Before Corrections in RL

PipelineRL uses vLLM as the inference engine for rollout generation. The
inference engine samples tokens and returns token logprobs; the trainer uses
those logprobs to compute policy ratios, KL, clip rate, entropy, and reward.
Any discrepancy in how those logprobs are computed can change the training
dynamics. This is the train-inference mismatch we needed to eliminate during
the vLLM V0 to V1 migration.

TL;DR. vLLM V1 matched our vLLM V0 reference after we fixed four things:
processed rollout logprobs, V1-specific runtime defaults, the inflight
weight-update path, and the fp32 lm_head used for the final projection. We
fixed the backend behavior before changing the RL objective.

The reference run used vLLM 0.8.5; the V1 runs used vLLM 0.18.1. Figure 1
shows the final result. The red run is the initial V1 attempt, and the green
run is the final V1 run after the fixes described below.

Migration Objective

vLLM V1 is a substantial rewrite of the V0 engine. Our migration target was
therefore deliberately narrow:

1. verify that V1 returned rollout logprobs in the form the trainer expected
2. rerun the same workload against the V0 reference
3. evaluate objective-level changes only after backend parity was restored

The first visible symptoms appeared in:

• clamp_log_ratio_new_old_indicator
• kl_new_old
• entropy
• reward

Those metrics came from a GSPO training run, the objective used for this
experiment. The same class of mismatch can surface in PPO, GRPO, or any online
RL system that treats rollout-side logprobs as part of the optimization target.

The initial V1 run showed the problem clearly. The trainer-side logprobs and
reward moved away from the V0 reference early in training.

The same pattern appears in the trainer metrics. Clip rate is the easiest signal
to read in the initial comparison.

Failure Modes

We separated the possible causes into three layers:

1. Semantic mismatch: the backend returns logprobs with different meaning
   relative to what the trainer expects.
2. Inference-path mismatch: the backend uses different runtime defaults for
   caching, scheduling, or request handling, so the same prompts follow a
   different execution path.
3. Objective mismatch: the RL objective needs correction for the amount of
   staleness or backend mismatch that remains.

We initially suspected the third category too early. The useful diagnosis came
from treating the first two as backend behavior problems and ruling them out
first.

V1 Backend Fixes

Logprob Semantics

The first issue was semantic. vLLM V1 returns logprobs from the raw model
outputs by default, before logits post-processing such as temperature scaling,
penalties, and top-k/top-p filtering. PipelineRL expected logprobs from the
processed distribution used by the sampler.

The required setting was:

• logprobs-mode=processed_logprobs

This removed the obvious mean offset in rollout logprobs. The training curves
still showed a gap relative to the known-good reference, so the next issue had
to be in the inference path.

The policy-ratio plot shows this directly. Once processed_logprobs is on for
V1, the mean policy ratio stays centered extremely close to 1.0 across all
three runs. That establishes the mean-bias fix. The remaining mismatch shows up
in clip rate, KL, entropy, and downstream training behavior.

Runtime Defaults

The early V1 run mixed the engine version with V1 runtime defaults:

• prefix caching, left unset in the early run so the vLLM 0.18.1 default
  applied
• async scheduling, left unset in the early run so the vLLM 0.18.1 default
  applied
• an ad-hoc disable-cascade-attn override that was set through launch-time
  kwarg passthrough and sits outside the parity recipe in committed config

For the parity run, we made these choices explicit:

vllm_config:
  use_v1: true
  vllm_kwargs:
    logprobs-mode: processed_logprobs
    enable-prefix-caching: false
    async-scheduling: false

Prefix caching deserves a separate note. It is normally a
correctness-preserving inference optimization for a fixed model state. In this
online RL setup, it was a V1-only difference in cache lifetime and reuse
relative to the V0 reference path. The actor was also handling repeated
prefixes, concurrent requests, async scheduling, and inflight weight updates.

A prefix-cache hit can reuse state computed before a weight update when the
cache policy ignores the weight-update boundary. Disabling prefix caching
removed one V1-only degree of freedom from the parity comparison.

Inflight Weight Updates

Weight synchronization also had to match the online-RL update model. One option
was to make V1 stricter than V0 by draining requests and clearing caches at
every update. That would answer a separate question. We first needed to verify
that V1 could match the existing V0 behavior.

What V0 effectively did was closer to:

• block execution at an engine boundary
• load the new weights
• resume without an explicit cached-state invalidation

The nearest V1 analogue was:

await engine.pause_generation(mode="keep", clear_cache=False)
await engine_client.collective_rpc_async(
    "receive_weight_update",
    args=(request.model_dump_json(),),
)
await engine.resume_generation()

Two details matter:

• mode="keep" matches the old inflight update model more closely than
  wait or abort
• clear_cache=False matches the V0 wrapper behavior, which left cached state
  intact on update

Lag was a useful runtime diagnostic. The initial V1 path carries more
persistent lag later in training than the corrected V1 run.

The Remaining Gap: fp32 lm_head

The V1 backend fixes above removed the obvious migration issues, but final
parity still required matching the numerical path used to compute logits. The
trainer used an fp32 lm_head for the final projection. The rollout backend
had to match that behavior.

A closely related issue appears in the
MiniMax-M1 technical report: their RL run
showed a training/inference token-probability mismatch that they traced to the
LM output head and fixed by computing the head in fp32.

This matters because the RL update consumes token logprobs directly. Small
changes in logits can become visible in policy ratios, KL, and clipping. The
final projection precision is therefore part of the correctness surface for
online RL. The
ScaleRL paper later includes fp32
logits/head computation as part of its RL recipe and ablates it as a useful
design choice for large-scale RL.

With the fp32 lm_head path included, reward gives a compact view of the final
parity result. In Figure 6, the final V1 run tracks the V0 reference; the
initial V1 attempt produces a clearly different reward curve.

Ablations

The negative results are important because they rule out common explanations.

• processed_logprobs alone: fixed the semantic logprob bug; the training
  mismatch remained.
• Batch invariance: the mismatch remained in a separate test, with higher
  lag, higher clip rate, and NCCL complications.
• Treating the first V1 run as a fair baseline: the first V1 run had
  multiple V1-only defaults enabled, so it was a confounded migration
  comparison.

Why We Fixed Backend Correctness First

Objective-side corrections such as truncated importance sampling,
importance-ratio reweighting, and related methods are useful tools. If rollouts
are intentionally stale, generated asynchronously, or produced by a backend
where equivalence to the trainer-side policy is unavailable, then some form of
correction is often the right thing to add.

The first problem here was inference correctness. After moving to V1, the
rollout backend returned logprobs and runtime behavior that broke the trainer
assumption. Adding an objective-side correction at that point would have mixed
two questions:

• is the inference backend producing the right logprobs?
• given correct logprobs, does the objective still need an off-policy or async
  correction?

Those questions need to be separated. Otherwise an objective-side correction
can compensate for broken inference-backend behavior, which makes the training
curve harder to interpret.

The current objective can still improve. After inference parity is restored,
the next improvement is the usual async/off-policy cleanup:

• keep explicit behavior-policy logprobs from rollout time
• recompute trainer-side old-policy logprobs at optimization time
• separate backend mismatch correction from the policy-update ratio
• track diagnostics like ESS for the correction term alongside aggregate
  trainer metrics

The main lesson from this migration is narrower: fix backend correctness first,
then add corrections for the mismatch that remains.