[grorge123]

Summary

Code-quality fixes from a static-analysis pass (GitRoll / clang-tidy). There are no API changes for callers.

1. Stop shelling out (fix: commit)

• fnv1aHash() (mlx/base.cpp) saved the array to ./temp_array.npy and ran popen("python3.10 ../hash_script.py ..."). hash_script.py doesn't exist in the repo, so debugArray() always threw. It now computes a 64-bit FNV-1a hash over the array's row-major bytes in C++, using mx::contiguous + eval(). There's no temp file, no Python, and no shell.
  • The old script is missing, so I can't confirm the new value matches what it returned. This is a debug helper, so I don't expect anything to depend on the exact value.
• example/vlm.cpp used std::system("python3.10 ../example/decode.py " + ModelPath + " output.npy"). It now runs the same command through posix_spawnp with an argv array. ModelPath is passed verbatim, so paths with spaces or shell characters work and nothing is interpreted by a shell. The interpreter (python3.10), script and output file are unchanged. If the spawn fails, it logs an error, as before.

2. const & for read-only parameters (refactor: commit)

About 45 mx::array / std::string / std::shared_ptr parameters that are only read now take const &. Header declarations and definitions are changed together.

I deliberately left these alone:

• Virtual functions: Embedding::forward, Linear::forward, Prompt::prepare. Their overrides (QuantizedEmbedding, QuantizedLinear, the *Prompt classes) would also have to change, so they're out of scope here.
• "pass by value and move" suggestions: these need std::move rather than const &, so they belong in a separate change.

Not changed

• bugprone-branch-clone: an earlier scan flagged 4 "identical if/else branches", in llm/transformer.cpp:104, llm/transformer.h:177, mllama/language.cpp:335 and vlm_base.cpp:351. All 4 are false positives. They came from a run that was missing the spdlog/fmt include paths. The branches really do differ: nn::RMSNorm vs the Gemma llm::RMSNorm, RopeScaling vs {}, and Cross vs Self decoder layers. With the real compile_commands.json, clang-tidy reports none of them.
• whisper.cpp integer division (Channels / 2 - 1, NState / NHead): this matches the reference Python's //, so it's intended.
• whisper_transcribe.cpp:46 still builds the ffmpeg command string and calls popen, so FilePath isn't quoted. Fixing that needs a spawn plus a pipe, which is a bigger change. I'd suggest a follow-up.

Verification

All of this ran on Linux x86_64 with clang 23 and MLX 0.32.4 (CPU wheel):

• Build: fresh CMake/Ninja builds before and after (-Wall -Wextra -Wpedantic -Werror). The same 29 project translation units compile in both. The ones that fail are the same before and after, and none are caused by this PR:
  • fmt and sentencepiece, which don't build with clang 23 / libstdc++ 15
  • gemma3.cpp, which hits a libstdc++ 15 std::optional constraint error. Its error output is identical before and after.
• clang-tidy with the real compile database: findings go from 135 → 54.
  • performance-unnecessary-value-param: 93 → 14 (the virtual and std::move cases above)
  • bugprone-command-processor: 3 → 1 (the ffmpeg one)
• fnv1aHash: I compiled the new function on its own against libmlx and compared it to a NumPy reference. It matches for a 3-D float32 array, a transposed non-contiguous view of it, and an int32 vector.
• spawn: the same posix_spawnp code received "../my models/x; echo INJECTED" as a single argv entry. A missing binary takes the error branch.

Not verified: macOS / Metal, linking and running the example binaries, and model inference. CI on this PR covers the Ubuntu and macOS builds.

🤖 Generated with Claude Code

[Copilot]

🟢 Approval recommended

The helper replacements are focused, the signature changes are aligned, and no blocking correctness issues were identified.

No open issues 🎉

What changed in this PR

This PR removes shell dependencies from two helpers and reduces unnecessary parameter copies across the MLX layers and model code.

Changes:

• Replaces Python-based array hashing with an in-process 64-bit FNV-1a implementation.
• Uses posix_spawnp to launch the VLM detokenizer without shell interpretation.
• Changes read-only parameters to const references, keeping declarations and definitions aligned and leaving virtual interfaces unchanged.

File	Description
model/​whisper/​decoding.h	Updates decoding declarations to const references.
model/​whisper/​decoding.cpp	Matches decoding definitions to updated declarations.
model/​whisper_transcribe.h	Updates model parameters in transcription declarations.
model/​whisper_transcribe.cpp	Matches transcription definitions to updated declarations.
model/​vlm_base.h	Updates the attention-mask input declaration.
model/​vlm_base.cpp	Passes the attention-mask input by const reference.
model/​mllama/​language.h	Updates attention and decoder cache declarations.
model/​mllama/​language.cpp	Passes cache pointers by const reference.
model/​llm/​transformer.h	Updates transformer array parameter declarations.
model/​llm/​transformer.cpp	Uses const references for arrays and helper strings.
model/​converter.h	Updates weight-path parameter declarations.
model/​converter.cpp	Passes weight paths by const reference.
mlx/​quantized.h	Updates quantization factory declarations.
mlx/​quantized.cpp	Passes source module pointers by const reference.
mlx/​positional_encoding.h	Updates the RoPE input declaration.
mlx/​positional_encoding.cpp	Passes RoPE input by const reference.
mlx/​normalization.h	Updates normalization input declarations.
mlx/​normalization.cpp	Passes normalization inputs by const reference.
mlx/​embedding.h	Updates the non-virtual asLinear declaration.
mlx/​embedding.cpp	Passes asLinear input by const reference.
mlx/​convolution.h	Updates convolution input declarations.
mlx/​convolution.cpp	Passes convolution inputs by const reference.
mlx/​base.h	Updates module registration and parameter interfaces.
mlx/​base.cpp	Implements native hashing and matching parameter changes.
mlx/​activations.h	Updates activation input declarations.
mlx/​activations.cpp	Passes activation inputs by const reference.
example/​vlm.cpp	Launches the detokenizer with explicit process arguments.

_{Review effort: Balanced}