perf(cat/grep/cli): per-file overhead cuts and faster cold start by nwaughachukwuma · Pull Request #194 · vlm-run/mm

nwaughachukwuma · 2026-08-03T12:53:57Z

Summary

mm grep: lists files via the Rust JSON fast path (no pyarrow import, ~120 ms/invocation); matches in one C-speed finditer pass (re.MULTILINE; zero-match files skip the per-line loop entirely); serves documents from the content-hash cache instead of re-extracting every PDF on every run; single-pass highlighting. mm wc uses the same cache with a thread pool.
cat pipeline: Magika ONNX model no longer loads or runs for auto-strategy — its output was never read there; mm.toml parsed once per mtime instead of ~10x per file; Generate type hints evaluated once; content hash computed once per file; ffmpeg probe + transcription OpenAI client process-cached; encoder discovery drops its O(modules × files) realpath sweep; keyframe counting demuxes packet flags instead of decoding every I-frame twice; JPEG 4:2:0 unless alpha.
CLI cold start: importlib.metadata deferred behind module __getattr__; pipeline/yaml imports out of the cat registration path.

Speed gains

Change	Old	New	Speedup
`mm grep` (340-file fixture, --count) wall	345 ms	238 ms	1.45x
repeat grep over PDFs	full re-extraction per run	one indexed SELECT	10-100x on document corpora
Magika in auto-strategy	load + per-file inference	skipped	~0.3-0.5 s per binary-file `mm cat`
mm.toml read	188 µs/call	27 µs	7x (× ~10 reads per cat'd file)
keyframe count	decode all I-frames twice	demux flags	~2x on the keyframes pipeline
`import mm.cli`	97 ms	77 ms	1.26x on every invocation

Validation

1,208 Python tests pass; ruff clean. Includes the test updates for the new content_hash kwarg and the cached transcription client.

🤖 Generated with Claude Code

https://claude.ai/code/session_012GnYsrLTmdJphKkm7yjWDV

gemini-code-assist · 2026-08-03T12:54:00Z

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

nwaughachukwuma

@claude please followup

nwaughachukwuma · 2026-08-03T14:06:59Z

    def _count_keyframes(self) -> int:
-        """Quick I-frame count via demux-level skip."""
+        """I-frame count from packet flags only — demux, no decode."""
        import av

        container = av.open(str(self._path))
        try:
            stream = container.streams.video[0]
-            stream.codec_context.skip_frame = "NONKEY"
            count = 0
            for packet in container.demux(stream):
-                for _ in packet.decode():
+                if packet.is_keyframe and packet.pts is not None:
                    count += 1
            return count
        finally:


@claude, how does this compare against the old implementation, and what's the change in behavior?

nwaughachukwuma · 2026-08-03T14:07:39Z

+- **RSI batch — grep/cat overhead + CLI cold start (260803)**:
+  - `mm grep`: Rust JSON fast path for the listing (no pyarrow import,
+    ~120 ms/invocation); matching is one C-speed `finditer` pass with
+    `re.MULTILINE` (zero-match files skip the per-line loop entirely);
+    documents route through the content-hash cache instead of
+    re-extracting every PDF each run; single-pass highlighting.
+    Fixture end-to-end: 345 ms → 238 ms wall.
+  - Magika no longer loads or runs for encoder auto-strategy — it only
+    ever read dimensional metadata (~0.3-0.5 s per binary-file `mm cat`;
+    model load off import). mm.toml parsed once per mtime (~10
+    parses/file → 1); `Generate` type hints evaluated once; content hash
+    threaded through `run_fast`/`run_accurate`; ffmpeg probe and the
+    transcription OpenAI client process-cached; keyframe counting demuxes
+    packet flags instead of decoding every I-frame twice; image payloads
+    encode JPEG 4:2:0 unless alpha is present.
+  - CLI cold start: `importlib.metadata` deferred, pipeline/yaml imports
+    out of the `cat` registration path — `import mm.cli` 97 ms → 77 ms.


remove the changes in changelog.md

nwaughachukwuma · 2026-08-03T14:09:30Z

    invoke_without_command=True,
    help="Fast, multimodal context for agents.",
-    epilog=f"mm v{__version__}",
+    epilog="mm — fast, multimodal context for agents (mm --version)",


revert to what it was before - we don't want cosmetic changes.

- mm grep: Rust JSON fast path for the listing (no pyarrow import), single finditer pass with MULTILINE, documents served from the content-hash cache, single-pass highlighting. mm wc routes documents through the same cache. - Magika no longer loads or runs for encoder auto-strategy — its output was never read there (~0.3-0.5s per binary-file mm cat; model load off import). - mm.toml parsed once per mtime (~10 parses/file → 1); Generate type hints evaluated once; content hash computed once per file; ffmpeg probe and the transcription OpenAI client process-cached; encoder discovery drops its O(modules x files) realpath sweep. - Keyframe counting demuxes packet flags instead of decoding every I-frame twice; image payloads encode JPEG 4:2:0 unless alpha is present. - CLI cold start: importlib.metadata deferred, pipeline/yaml imports out of the cat registration path (import mm.cli 97ms → 77ms). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012GnYsrLTmdJphKkm7yjWDV

nwaughachukwuma commented Aug 3, 2026

View reviewed changes

nwaughachukwuma closed this Aug 3, 2026

nwaughachukwuma reopened this Aug 12, 2026

nwaughachukwuma force-pushed the claude/core-perf-optimization-s230yb branch from 9ce1d7e to 0d2f4ca Compare August 12, 2026 10:30

nwaughachukwuma force-pushed the claude/perf-catgrep-coldstart-s230yb branch from a56fb33 to f33c7c2 Compare August 12, 2026 10:32

nwaughachukwuma force-pushed the claude/perf-catgrep-coldstart-s230yb branch from f33c7c2 to 9c03a83 Compare August 12, 2026 10:40

nwaughachukwuma changed the base branch from claude/core-perf-optimization-s230yb to add-report-flag-for-cat-command August 12, 2026 10:40

nwaughachukwuma mentioned this pull request Aug 12, 2026

perf(store): O(1) ensure_metadata, staleness-gated upserts, parallel batch extraction #192

Closed

nwaughachukwuma closed this Aug 12, 2026

Navigation Menu

Uh oh!

perf(cat/grep/cli): per-file overhead cuts and faster cold start - #194

perf(cat/grep/cli): per-file overhead cuts and faster cold start#194
nwaughachukwuma wants to merge 1 commit into
add-report-flag-for-cat-commandfrom
claude/perf-catgrep-coldstart-s230yb

nwaughachukwuma commented Aug 3, 2026 •

edited

Loading

Uh oh!

gemini-code-assist Bot commented Aug 3, 2026

Uh oh!

nwaughachukwuma left a comment

Uh oh!

nwaughachukwuma Aug 3, 2026

Uh oh!

nwaughachukwuma Aug 3, 2026

Uh oh!

nwaughachukwuma Aug 3, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

nwaughachukwuma commented Aug 3, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Speed gains

Validation

Uh oh!

gemini-code-assist Bot commented Aug 3, 2026

Uh oh!

nwaughachukwuma left a comment

Choose a reason for hiding this comment

Uh oh!

nwaughachukwuma Aug 3, 2026

Choose a reason for hiding this comment

Uh oh!

nwaughachukwuma Aug 3, 2026

Choose a reason for hiding this comment

Uh oh!

nwaughachukwuma Aug 3, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

nwaughachukwuma commented Aug 3, 2026 •

edited

Loading