Add Qwen3.5 model support (text-only) and fix reasoning+tool streaming

[otarkhan]

Summary

• Qwen3.5 text-only loading: Adds a strict=False fallback in the tokenizer loader for models whose weight files contain extra parameters (e.g., vision tower weights). When mlx-lm rejects a model with "parameters not in model", the loader retries with strict=False to discard the unrecognized weights and load the model as text-only.
• Fix streaming tool calls with reasoning parser: The streaming code path had mutually exclusive branches for reasoning parsing and tool call parsing. When both --reasoning-parser and --tool-call-parser were enabled simultaneously, tool calls were never parsed during streaming — the raw XML markup (e.g., <tool_call><function=bash>...</function></tool_call>) leaked through as content. Now the reasoning parser's content output is piped through the tool parser before being emitted.
• Dynamic memory pressure threshold: The emergency memory pressure threshold was hardcoded at 200GB, which caused constant forced cache clears on high-memory systems running large models. It now dynamically queries Metal's max_recommended_working_set_size and sets the threshold to 85% of that value, scaling correctly with system RAM.

Why text-only for Qwen3.5?

Qwen3.5 models (e.g., Qwen3.5-397B-A17B) are vision-text models that include a vision tower (SigLIP encoder). Loading them through the MLLM path (mlx-vlm) fails because:

1. mlx-vlm does not support the qwen3_5_moe architecture yet
2. Even if it did, the MoE layers produce ArraysCache instead of KVCache, which is incompatible with the MLLM batch generator's continuous batching (slicing, merging, and cache management all assume KVCache)

Loading via mlx-lm with strict=False discards the ~333 vision tower parameters and serves the model as text-only. This works well with continuous batching, prefix caching, reasoning, and tool calling.

Files changed

• vllm_mlx/utils/tokenizer.py — Add _load_strict_false() and fallback for "parameters not in model" error
• vllm_mlx/server.py — Integrate tool call parsing into the reasoning parser streaming branch
• vllm_mlx/engine_core.py — Dynamic memory pressure threshold based on Metal device info

Testing

Manually tested with:

vllm-mlx serve mlx-community/Qwen3.5-397B-A17B-4bit \
    --host 0.0.0.0 \
    --port 8081 \
    --max-num-seqs 5 \
    --cache-memory-percent 0.2 \
    --enable-prefix-cache \
    --reasoning-parser qwen3 \
    --enable-auto-tool-choice \
    --tool-call-parser qwen3_coder \
    --continuous-batching \
    --max-tokens 262144 \
    --served-model-name Qwen3.5-397B-A17B

Verified:

• Model loads successfully with vision tower weights discarded
• Reasoning (<think> blocks) extracted correctly
• Tool calls parsed correctly during streaming (multi-turn tool use works)
• No memory pressure warnings on 512GB unified memory system
• Prefix cache functional with continuous batching

[waybarrios]

thanks for this, the tokenizer strict=False fallback and the dynamic memory threshold are solid changes

i made two small adjustments and pushed them to your branch

1. removed the server.py streaming changes since PR fix: integrate tool call parsing with reasoning parser in streaming mode #148 already covers the reasoning+tool parsing fix with a slightly different approach that handles more edge cases (tool calls inside reasoning blocks, transition chunks). keeping both would conflict so i dropped yours in favor of fix: integrate tool call parsing with reasoning parser in streaming mode #148

2. fixed a bug in _load_strict_false where the model config was being discarded. load_model returns (model, config) and the config has the eos_token_id that needs to be passed to load_tokenizer, otherwise the model might not stop generating at the right token. the fix looks like this

model, config = load_model(model_path, strict=False)
tokenizer = load_tokenizer(
    model_path,
    tokenizer_config or {},
    eos_token_ids=config.get("eos_token_id", None),
)

also removed the redundant from pathlib import Path since its already imported at module level

the rest of the PR (tokenizer fallback + dynamic memory pressure) looks good to merge. let me know if you have any questions