docs+cli: sync imp.conf.example, --help text and config comments with parser reality#878
Merged
Merged
Conversation
… parser reality Three-way consistency sweep (RuntimeConfig parser vs imp.conf.example vs CLI parsers vs docs). No hot-path behavior change. - imp.conf.example: add the 4 parsed-but-undocumented keys (runtime.prefill_chunk_decode_cap, attention.mla_absorb, gemm.nvfp4_lm_head_cutlass, diagnostics.spec_capture_probe) — all 132 parsed keys are now covered; fix 'imp --set' -> 'imp-cli --set'. - config.h: fix stale comments (q8_imma_enabled / moe_imma_prefill say 'default off' but default true; fp8_auto_legacy claimed an IMP_KV_FP8_AUTO env that seed_from_env never reads) and the matching phantom-env log line in engine_init_resolver.cpp. - imp-cli --help: add --config, --set, --perplexity, --repeat-last-n (all parsed + functional but undocumented). - imp-server --help: add --config, --set, --min-kv-tokens, --prefix-cache; align --prefill-chunk-size default wording with the CLI (-1 = per-arch, not 0). - Remove the dead turboquant_sketch_mult/multiplier fields from both args.h and EngineConfig (--tq-sketch-mult stays parsed-and-ignored; ImpConfig keeps its field for ABI compat). - imp-bench: 'all' now actually runs the nvfp4 benchmark; usage text documents that the no-arg default excludes it. - docs/usage.md: add the missing server flags (--host, --port, --max-batch, --models-dir, --lora, --request-timeout, --max-input-tokens, --prefix-cache) to the server-flags table. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
kekzl
added a commit
that referenced
this pull request
Jul 7, 2026
…liases, dead flags (#879) Simplification pass over the config/CLI layer (follow-up to #878). Every removal was verified to have zero producers/consumers in the repo; all functionality stays reachable via imp.conf / --set. - seed_from_env: 29 legacy IMP_* env reads -> 2. Only IMP_DETERMINISTIC (set by test_determinism_e2e/test_lora) and IMP_FMHA_FA2 (set by the roofline A/B harness) have real producers; the other 27 had none. Stale 'Legacy env: IMP_*' comments and log strings swept repo-wide. - Fix --kv-fp16: it set an IMP_KV_FP16 env var that nothing has ever read (silent no-op). It now pushes kv_cache.dtype=fp16, actually forcing FP16 and opting out of the auto FP8-hint upgrade. - --no-fp8-prefill: setenv round-trip replaced by a direct attention.fp8_prefill=never config override. - Retire deprecated TurboQuant surface: --kv-turboquant{,-lite} and --tq-sketch-mult flags (CLI + server + per-request JSON overrides), one-shot WARN blocks, IMP_DTYPE_TURBOQUANT(_LITE) C-API aliases, and the ignored turboquant_sketch_multiplier field. TurboQuant was retired Phase 5 (2026-05-17); the aliases had no callers. - Delete the --cuda-graphs no-op flag (graphs are default-on). - Delete dead ImpConfig fields gpu_memory_pool_size and num_cpu_threads (default-initialized, never read). Verified: make build + test-unit (37/37) + verify-fast (fast gtest PASS); GPU smoke on Qwen3-8B-Q8_0 — --kv-fp16/--no-fp8-prefill resolve correctly (log: 'FP8 prefill: disabled (attention.fp8_prefill=never)'), generation coherent; removed flags now fail fast with 'Unknown argument'. Co-authored-by: Claude Fable 5 <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.
What
Three-way consistency audit of the config surface —
RuntimeConfigparser (src/runtime/config.{h,cpp}) vsimp.conf.examplevs theimp-cli/imp-serverargument parsers vsdocs/usage.md— and the fixes for every verified drift point. No hot-path behavior change.Findings & fixes
imp.conf.example (billed as the full schema by docs/usage.md) was missing 4 parsed, live keys:
runtime.prefill_chunk_decode_cap,attention.mla_absorb,gemm.nvfp4_lm_head_cutlass,diagnostics.spec_capture_probe— added with intent comments. All 132 parsed keys are now covered (verified by extracting every key string from config.cpp and grepping the example). Alsoimp --set→imp-cli --setin the header.Stale comments in config.h (code right, comment wrong):
q8_imma_enabledandmoe_imma_prefillsaid "default off" — defaults aretrue.fp8_auto_legacyclaimedIMP_KV_FP8_AUTOenv compat, butseed_from_envnever reads that env; comment now says imp.conf-key-only. The matching phantom-env log line inengine_init_resolver.cppfixed too.--help gaps (flags parsed + functional but absent from usage text):
--config,--set,--perplexity,--repeat-last-nadded.--config,--set,--min-kv-tokens,--prefix-cacheadded;--prefill-chunk-sizewording aligned with the CLI (default is -1 = per-arch, not 0).Dead code:
turboquant_sketch_mult(bothargs.h) andEngineConfig::turboquant_sketch_multiplierwere written-never-read — removed.--tq-sketch-multstays parsed-and-ignored andImpConfigkeeps its field for ABI compat.imp-bench:
alldid not run thenvfp4benchmark — now it does; usage text documents that the no-arg default excludes it (isolated ncu target).docs/usage.md: server-flags table was missing
--host,--port,--max-batch,--models-dir,--lora,--request-timeout,--max-input-tokens,--prefix-cache— added.Verification
make build+make test-unit(37/37) +make verify-fast(fast gtest filter PASS; model-gated perf/smoke stages skipped — models not mounted in the verify container; not needed, no hot-path change).--helpoutput checked in the built image;imp-cli --config imp.conf.exampleparses with zero unknown-key warnings.section.keyparsed in config.cpp appears in imp.conf.example.No perf impact; baseline untouched.
🤖 Generated with Claude Code