Conversation
This was referenced Jun 1, 2026
aoshen02
added a commit
to vllm-project/vime
that referenced
this pull request
Jun 5, 2026
…#1940) Port of THUDM/slime#1939 (add more cpu ci) + #1940 (run cpu test on main): - tests/test_rm_{deepscaler,f1,gpqa,math,math_dapo}.py: pure-cpu unit tests for vime/rollout/rm_hub reward fns (107 tests, validated locally). - vime/utils/misc.py: lazy 'import ray' so cpu-only paths (rm_hub scoring, plugin contracts) don't require ray. - pr-test.yml.j2: register the 5 rm tests in the cpu (num_gpus:0) matrix; enable push-to-main trigger gated to cpu jobs only (GPU/self-hosted jobs stay PR/label-gated so push never burns the fleet). slime #1939 also added tests/test_sample.py, which references Sample.rollout_id (introduced by the #1926-chain / vime PR #119) — that test is stacked on #119 instead of here. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Signed-off-by: aoshen02 <aoshen@inferact.ai>
CalvinXKY
pushed a commit
to vllm-project/vime
that referenced
this pull request
Jun 7, 2026
…80/#1967/#1938/#1988) (#138) * [CI] add reward-fn cpu tests + cpu-on-main trigger (port slime #1939+#1940) Port of THUDM/slime#1939 (add more cpu ci) + #1940 (run cpu test on main): - tests/test_rm_{deepscaler,f1,gpqa,math,math_dapo}.py: pure-cpu unit tests for vime/rollout/rm_hub reward fns (107 tests, validated locally). - vime/utils/misc.py: lazy 'import ray' so cpu-only paths (rm_hub scoring, plugin contracts) don't require ray. - pr-test.yml.j2: register the 5 rm tests in the cpu (num_gpus:0) matrix; enable push-to-main trigger gated to cpu jobs only (GPU/self-hosted jobs stay PR/label-gated so push never burns the fleet). slime #1939 also added tests/test_sample.py, which references Sample.rollout_id (introduced by the #1926-chain / vime PR #119) — that test is stacked on #119 instead of here. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Signed-off-by: aoshen02 <aoshen@inferact.ai> * [Doc] sync customization/trace/pd-disagg docs (port slime #1942) Port of THUDM/slime#1942 (update docs). Brought the technical doc updates (customization.md agentic-workflows section, trace.md, pd-disaggregation.md, en+zh), rebranded slime->vime / sglang_rollout->vllm_rollout / SGLang->vLLM. Divergence handling: - README.md / README_zh.md: kept vime's version (slime's changes were SGLang-branded architecture text + 'Projects built upon slime' marketing + an agentic-examples list — vime uses vLLM and trimmed examples in #126). - Dropped the broken examples/search-r1 link from the added customization text (trimmed in vime); kept examples/multi_agent + examples/fully_async (exist). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Signed-off-by: aoshen02 <aoshen@inferact.ai> * [Chore] remove redundant replay tool + harden unwrap_model import (port slime #1943) Port of THUDM/slime#1943 (remove redundant file): - remove tools/replay_openai_jsonl.py (redundant). - vime/backends/megatron_utils/model.py: make unwrap_model import resilient to the Megatron move from megatron.core.utils to megatron.core.pipeline_parallel.utils (try/except fallback). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Signed-off-by: aoshen02 <aoshen@inferact.ai> * [Sync][C] FLOPs MLA fix, PYTHONUNBUFFERED typo, spec-config guard, fault-tolerance docs Mechanical / docs sweep for the slime→vime sync (mega-PR C additions, on top of the cherry-picked #1939+#1940 / #1942 / #1943): - slime #1980: fix FLOPs accounting for non-MLA attention. Gate the MLA-shaped qkv/attention flop formulas on `multi_latent_attention` instead of inferring MLA from q_lora_rank/qk_pos_emb_head_dim/v_head_dim (which misattributed flops for non-MLA models). vime/utils/flops_utils.py. - slime #1967: fix PYTHONBUFFERED=16 typo -> PYTHONUNBUFFERED=1 across 12 scripts + command_utils.py (both name and value, matching slime canonical). - slime #1938: guard `vllm_speculative_config` read in _compute_spec_metrics with getattr(...) so --debug-train-only (skip_vllm leaves the attr unset) doesn't AttributeError. vime/ray/rollout.py. - slime #1988: rewrite docs/en/advanced/fault-tolerance.md (scope, health checks, debug/replay path, production pattern), rebranded sglang→vllm / slime→vime, /health_generate→/health, link vllm-config.md. Dropped from this PR after source review: - #1987/#1990 (R3 ci ref/logprob+kl guards) — already in main via vime #93 (b1009dd "Backports slime #1987 + #1990"). - #1974/#1989 — touch examples/coding_agent_rl + its README bullet; that feature is not in vime yet (lands with the agent mega-PR). - #1975 — sglang conda-ci image resolution + a test vime already has + build_conda micromamba fix; no real version-file port for vime. Deferred to after the rollout-data-model PR: #1986/#1985 (they edit that PR's tests). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Signed-off-by: aoshen02 <aoshen@inferact.ai> * docs(mega-C): faithfully port slime #1942 docs (fault-tolerance/pd-disagg/trace) Audit (一模一样): mega-C's original #1942 doc port was incomplete — several docs were stubs or truncated vs slime. Port them faithfully (modulo sglang→vllm): - fault-tolerance.md en: add trailing (trainer-failure note + Related Docs). - fault-tolerance.md zh: was a 13-line stub → full 76-line port. - pd-disaggregation.md en+zh: were 7-line stubs → full 87-line port. - trace.md en+zh: port slime's build_meta example as build_vllm_meta_trace_attrs(output) (vime's real signature). Translation: SGLang→vLLM, /health_generate→/health (vime's real endpoint), sglang-config→vllm-config, `sglang:`→`vllm:` YAML key (vime vllm_config.py requires the 'vllm' key), server→engine; `--prefill-num-servers` flag preserved. All 6 files now match slime line counts (76/76, 87/87, 119/119). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * docs(mega-C): port slime #1942 customization gaps (multi-sample return + teacher_log_probs) Surgically add the two slime #1942 additions that mega-C missed, keeping vime's legitimate divergences (vllm_rollout path, VIME_CONTRACT_* env, search-r1 removed in #74): - "#### Returning multiple training samples for one prompt" section (en+zh): custom_generate may return list[Sample] with shared group_id; signature updated to `-> Sample | list[Sample]`. - `teacher_log_probs` field in the Sample-fields dict (en+zh). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * [Sync][C] align to slime exactly: drop #1961 over-port, fix #1938/#1943 placement - customization.md (en+zh): revert custom_generate signature to `-> Sample` and remove the "Returning multiple training samples" section. That content is from slime #1961 (not #1942, despite the prior commit message) and documents Sample.group_id / Sample.rollout_id + list[Sample] return — all tied to the deferred rollout-data-model refactor that vime does not yet implement (verified: 0 occurrences of group_id/rollout_id in types.py and the rollout loop). Defers with #1986/#1985. - model.py: move the unwrap_model try/except to slime #1943's exact position (after the tqdm import) — import block now byte-identical to slime. - rollout.py: drop the 4-line comment around the spec-metrics getattr guard; slime #1938 is a pure one-line change with no added comment. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * [Sync][C] cleanup sglang/dead-link residue: /health_generate help text + search-r1 refs - arguments.py: --rollout-health-check-{interval,timeout} help text said `/health_generate` (sglang's endpoint) but the code hits `/health` (vllm_engine.py). Fix the help strings to `/health`. - customization.md (en+zh) + index.rst (en+zh): remove the dead `examples/search-r1` references — that example does not exist in vime. Note: the method name `VLLMEngine.health_generate()` (vllm_engine.py) and its caller (health_monitor.py) keep the sglang-flavored name but already hit `/health`; renaming the method is a separate change, left out here. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * [Sync][C] remove dead examples/retool toctree refs examples/retool does not exist in vime (same dead-link situation as search-r1); drop it from the docs toctree in index.rst (en+zh). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * [Sync][C] docs: add Ray Distributed Debugger step-by-step section (port slime #2001) Appends the 'Step-by-Step Debugging with Ray Distributed Debugger' section (debugpy + RAY_DEBUG_POSTMORTEM + VS Code attach) to developer_guide/debug.md (en + zh). Content is engine-neutral (Ray/debugpy/VS Code), ported verbatim; no sglang->vllm translation needed. --------- Signed-off-by: aoshen02 <aoshen@inferact.ai> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.