[Kh4L]

Summary

• size completions against the fully rendered prompt when an exact tokenizer is available, while keeping approximate fallback estimates from causing false context-overflow failures
• add a configurable completion floor and set GDPVal to 8,192 tokens so long tasks do not collapse into truncated 1,024-token tool calls
• keep model context capacity separate from the Responses API max_output_tokens; request output limits now cap completion output without shrinking the assumed context window
• mirror Nemotron historical-reasoning truncation in estimator-only fallback mode without mutating the retained trajectory or provider request
• recognize length-truncated schema-invalid tool calls and apply a one-turn recovery instead of allowing repeated empty code calls
• sanitize leading inline reasoning from compaction summaries, retry empty or undersized summaries up to three times, and fail closed instead of replacing useful history
• keep generic Stirrup defaults compatible; GDPVal opts into a 262,144-token context window, an 8,192-token completion floor, and a 50-word compaction minimum

Motivation

A paired AA/Gym rollout analysis found two harness-level long-context failures. Raw-history fallback counting included reasoning that the served Nemotron template omitted, collapsing later completion budgets to 1,024 tokens and producing repeated invalid tool calls. Separately, reasoning-model compaction responses were replayed with their leading think blocks, adding millions of unnecessary prefill tokens and occasionally replacing progress with near-empty summaries.

Validation

At HEAD 470a5e324:

• Stirrup component suite: 286 passed
• focused budgeting, compaction, configuration, and output/context boundary coverage: 115 passed
• vLLM request-template boundary tests: 3 passed
• full pre-commit run --all-files: passed
• pinned iter-12500 tokenizer replay: exact template count 321, direct render count 321
• offline historical replay: the two budget-regression tasks retain a 64K budget across all 250 recorded prefixes; 94 historical compaction outputs sanitize to zero leaked leading think blocks, with the eight undersized summaries rejected
• explicit boundary regression: an 8K request output cap remains independent from the 262K model context window, including when the exact prompt is already larger than 8K

HSG real-checkpoint canary for the core budgeting/compaction commit 3e496e220:

• Slurm job 6924145: COMPLETED, exit 0, 4 GPUs
• checkpoint: /lustre/fs1/portfolios/llmservice/projects/llmservice_modelalignment_ppo/users/abukharin/super35_dualds0821_official_hf_pipeline/hf_full/full_generalist_prod/iter_12500
• 3/3 selected long-context tasks completed; 230/230 policy calls; zero rollout or transport failures
• completion budget min/max: 64K/64K; zero 1,024-token or sub-8K requests
• two natural 70%-threshold compactions were sanitized, replayed, and completed successfully (385- and 567-word retained summaries)
• fallback live probe reduced the false raw estimate from 260,270 to 270 tokens and dispatched a 64K completion
• compaction live probe stripped inline reasoning, retained a 209-word summary, and replayed it successfully
• immutable receipt: /lustre/fs1/portfolios/llmservice/projects/llmservice_nemotron_ultra/users/spanev/gdpval_context_fix_validation_20260904/results/job-6924145/context-fix-canary-receipt.json

HSG full GDPVal task-only validation at HEAD 470a5e324:

• exact source-of-truth set: 220/220 canonical rows, 0 failure rows, no duplicate or unexpected task IDs
• exact vLLM access-log accounting: 15,743/15,743 chat-completion requests returned 200
• 15,640 persisted policy turns plus 103 internal summary requests; the latter reconcile exactly to 99 accepted compactions and four rejected first attempts that were retried successfully
• all 99 retained compaction summaries had leading reasoning removed and met the 50-word floor; zero terminal compaction failures
• every persisted policy-call prompt left enough measured context headroom for the configured completion floor (maximum prompt: 193,343 tokens); 6,972 Ray-visible budget samples were all 64K, with zero 1,024-token or sub-8K samples
• empty code_exec({}) calls fell from 364 in the historical run to 1; leading-think retained summaries fell from 94 to 0; under-50-word summaries fell from 8 to 0
• 220 finish-marker files, 202 non-null finishes, 18 max-turn/null finishes; 19 tasks reached 250 turns
• aggregate concurrency 160 on 32 GPUs; collection wall time 2:55:37; independently scheduled shards consumed 61.4 GPU-hours
• finalizer job 6927667: COMPLETED, exit 0, all eight central audits PASS
• canonical rollout SHA-256: 2281eae7c1531f93181ab15dc5dc620f8aa8133bf8077cdd449f5fbf5e3ac966

One shard's campaign-local end-audit helper exited after collection because it used str.splitlines() on a valid JSON string containing U+2028. Its 28/28 LF-delimited rows were intact, and the corrected central audit passed them before canonical publication. This was isolated to the validation helper, not the Gym changes.

The historical comparison used the same ordered 220 tasks, equivalent template semantics, 32 GPUs, and aggregate concurrency 160. Its checkpoint was the pure_lm_ab/hf_text wrapper, while the full validation used official_hf_pipeline/hf_full; their language tensors map one-to-one, but the comparison is directional E2E evidence rather than a code-only causal A/B.

The follow-up HEAD commit's output-limit/context-capacity separation is causally covered by the dedicated 8K/262K regression tests; the full run used an outer limit of 262K and provides outcome evidence for that boundary.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.