[ganisback]

Summary

• After a model converted from SafeTensors loads in llama-server, delete the model*.safetensors shards and model.safetensors.index.json the converter consumed, so the checkpoint is not stored twice (e.g. JEV-9B kept ~18 GB of shards next to its 9.5 GB GGUF).
• Companion files such as head.safetensors and adapter directories are kept; a vision model keeps its weights until its mmproj GGUF exists.
• The manifest is refreshed (files, size, format) so cluster sync ships the GGUF as a regular file instead of an optional derived extra.
• Folded the two duplicated "start engine from a converted GGUF" blocks in loadEngineWithProgressMode into one closure.

Trade-off: converting the same model to a different dtype later needs a fresh download.

Testing

• Unit tests for the cleanup rules and manifest refresh; make lint; go test ./internal/{convert,model,inference,server}.
• End to end on a copy of laya-typed-decisions without its GGUF: conversion, load, 843 MB of SafeTensors removed, manifest switched to gguf, and a reload served the GGUF.

Review fixes (8663096)

• Removed files are recorded in manifest.json as removed_source_weights before deletion. A dtype request with no matching GGUF then fails with ErrSourceWeightsRemoved instead of serving the GGUF that is left.
• Leftover .safetensors companions no longer make the model look convertible. Removal waits for any conversion of the same directory, using a per-directory RWMutex.
• go test ./... passes, and the new tests also pass under -race.

End to end: Qwen/Qwen3.5-4B (vision, 2 shards, 8.7 GB) on the :11435 preview

Step	Result
POST /api/load with dtype=f16	Converted Qwen3.5-4B-f16.gguf (8.7 GB) and mmproj-Qwen3.5-4b-F16.gguf in ~31s, loaded ready. A chat completion answered correctly.
Source cleanup	Ran only after the mmproj existed. Removed 2 shards + index and freed 9.3 GB. The manifest now has format=gguf, removed_source_weights lists the 3 files, and only the 2 GGUFs remain.
POST /api/load with dtype=q8_0	HTTP 400: model Qwen/Qwen3.5-4B has no q8_0 GGUF: SafeTensors source weights were removed after an earlier GGUF conversion; pull the model again to convert another dtype. No conversion ran and the f16 GGUF was not served.
PUT /api/models/Qwen/Qwen3.5-4B/config	dtype=q8_0 was rejected with 400 (no q8_0 build and no weights to convert into one). dtype=f16 was accepted with 200. A default load afterwards was ready.

No panics and no data race reports in the server log.

🤖 Generated with Claude Code

[ganisback]

Three blocking issues with deleting the SafeTensors source after a converted GGUF loads. See the inline comments.

[ganisback]

Addressed the three blocking review comments in 8663096:

• dtype ignored after removal: the deleted files are now recorded in manifest.json (removed_source_weights) before they are removed. A dtype request with no matching GGUF now fails with ErrSourceWeightsRemoved instead of serving the GGUF that is left. A re-pull that restores every recorded file makes the model convertible again.
• companion .safetensors: HasConvertibleHFWeights returns false once the recorded shards are missing, so head.safetensors no longer looks convertible. Convert also refuses to run in that case.
• race: there is now a per-model-dir RWMutex. Convert holds the read lock and removal takes the write lock, so shards are never deleted while a conversion is reading them. Every load runs in the server process, so a lock inside that process is enough.

Tests: go test ./... passes, and the new tests also pass under -race.