fix(server): fix SMT multimodal crash on hybrid-recurrent models + drop temp-file round-trip Two related fixes for the SMT vision multimodal server path

[co-seven]

Two related fixes for the SMT vision multimodal server path:

1. Prompt-cache reuse crash on hybrid/recurrent multimodal models (e.g. Qwen3.5VL). A multi-turn chat answered turn 1 but aborted on turn 2 with: common.cpp:1495: failed to remove sequence N with p0=M, p1=-1 Root cause: the text prompt-cache reuse path (LCP-similarity slot selection + get_common_prefix) was applied to multimodal requests on hybrid-recurrent models. Their llama_memory_recurrent::seq_rm() cannot partially remove a tail range outside the bounded per-token snapshot window (n_rs_seq), so the reuse-computed KV position makes seq_rm return false -> GGML_ABORT. Pure-attention multimodal models (e.g. LLaVA-style fastvlm) are unaffected. Gate prompt-cache reuse off precisely for multimodal AND hybrid/recurrent models (prompt_cache_reuse_unsafe()), so those prompts are fully re-processed while pure-attention multimodal models keep their cache reuse.

2. Remove the pointless temp-file round-trip in SMT vision encode. The preprocessed float32 tensor was written to /tmp/llama-server-smt-XXXXXX and immediately read back into memory by the ONNX engine. Feed the in-memory blob directly (encode_image_mem / set_input_tensor_from_memory). This removes a per-image disk write+read and fixes intermittent "failed to write temp file" failures on boards whose root (/tmp) partition is full. The file-path API is kept as a thin wrapper for other callers; the audio path is unchanged.

[Copilot]

Pull request overview

Fixes two issues in the SMT multimodal server path: (1) avoids a crash when prompt-cache reuse interacts with hybrid/recurrent multimodal models, and (2) removes an unnecessary disk temp-file round-trip when feeding SMT vision tensors into the ONNX runtime.

Changes:

• Disable prompt-cache reuse (LCP-based slot selection + common-prefix reuse) for hybrid/recurrent multimodal models to prevent seq rollback aborts.
• Add an in-memory SMT vision encode path and switch the server to feed preprocessed tensor bytes directly to ONNX (no temp file).
• Bump version to 0.1.6.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
VERSION_NUMBER	Version bump to 0.1.6.
tools/server/server-smt-vision.cpp	Uses in-memory SMT vision encoding instead of writing/reading a temp file.
tools/server/server-context.cpp	Gates prompt-cache reuse off for hybrid/recurrent multimodal models to avoid recurrent-state rollback aborts.
tools/mtmd/smt-vision-wrapper.h	Exposes encode_image_mem() API for in-memory tensor ingestion.
tools/mtmd/smt-vision-wrapper.cpp	Implements in-memory tensor ingestion and keeps a file-path wrapper delegating to it.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 5 out of 5 changed files in this pull request and generated 1 comment.

[Copilot]

Pull request overview

Copilot reviewed 5 out of 5 changed files in this pull request and generated no new comments.