Update patches to match upstream PR jundot/omlx#1881

patches/README.md

@@ -1,27 +1,28 @@
 # oMLX patches for serving MiniMax-M3-oQ4
 
-These are the oMLX-side changes needed to **serve** this checkpoint (beyond mlx-vlm PR #1374). They
-are **not yet in oMLX main** — they apply cleanly onto current main (verified). A PR is being opened
-upstream; until merged, apply them to your oMLX install.
+These are the oMLX-side changes needed to **serve** this checkpoint (beyond mlx-vlm PR
+[#1374](https://github.com/Blaizzy/mlx-vlm/pull/1374)). They are **submitted upstream as
+[jundot/omlx#1881](https://github.com/jundot/omlx/pull/1881)** — once that merges, no patches are
+needed. Until then, they apply cleanly onto oMLX `main`.
 
-Apply from the oMLX package root (the dir containing `scheduler.py`, `server.py`, `oq.py`, `api/`):
+Apply from the `omlx/` package root (the dir containing `scheduler.py`, `server.py`, `oq.py`, `api/`):
 
 ```bash
-patch -p0 oq.py              < patches/oq.py.patch
-patch -p0 scheduler.py       < patches/scheduler.py.patch
-patch -p0 api/utils.py       < patches/api_utils.py.patch
+patch -p0 oq.py               < patches/oq.py.patch
+patch -p0 scheduler.py        < patches/scheduler.py.patch
+patch -p0 api/thinking.py     < patches/api_thinking.py.patch
 patch -p0 api/tool_calling.py < patches/api_tool_calling.py.patch
-patch -p0 server.py          < patches/server.py.patch
+patch -p0 server.py           < patches/server.py.patch
 ```
 
 | Patch | Purpose | Needed for |
 |-------|---------|-----------|
 | `scheduler.py.patch` | Serialize `minimax_m3_vl` (like Llama-4) + MiniMax-Sparse-Attention KV cache (`MiniMaxM3KVCache` ↔ batch variant) | running under oMLX batching / concurrency |
-| `api_utils.py.patch` | Map `<mm:think>`/`</mm:think>` → `<think>`/`</think>` | clean `reasoning_content` |
+| `api_thinking.py.patch` | Normalize `<mm:think>`/`</mm:think>` → `<think>`/`</think>` in `extract_thinking` + `ThinkingParser.feed` | clean `reasoning_content` |
 | `api_tool_calling.py.patch` + `server.py.patch` | Parse `<invoke name=…>` + bare `<key>value</key>` params; strip the `]<]minimax[>[` token (200058) from streamed content | structured `tool_calls`, no markup leak |
 | `oq.py.patch` | `_vlm_sanitize` proxy so oQ can quantize a VLM without instantiating it | **quantizing only** — not needed to serve this checkpoint |
 
 Notes:
-- The tool-call parser belongs more naturally in mlx-vlm's `tool_parsers` (then selectable via
-  `tokenizer_config.json` `tool_parser_type` with no oMLX patch); a separate mlx-vlm PR is the long-term home.
 - `trust_remote_code: true` and `torch`+`torchvision` in the serving env are also required (see the main README).
+- The tool-call parser could instead live in mlx-vlm's `tool_parsers` (selectable via `tokenizer_config.json`
+  `tool_parser_type`); see the discussion on #1881.

patches/api_thinking.py.patch

@@ -0,0 +1,29 @@
+diff --git a/api/thinking.py b/api/thinking.py
+index 024c0df..e245c14 100644
+--- a/api/thinking.py
++++ b/api/thinking.py
+@@ -168,6 +168,12 @@ def extract_thinking(text: str) -> Tuple[str, str]:
+     if not text:
+         return ("", "")
+ 
++    # MiniMax-M3 (minimax_m3_vl) emits reasoning wrapped in <mm:think>...</mm:think>;
++    # normalize to the standard <think>/</think> so it separates into reasoning_content.
++    # No-op when absent.
++    if "<mm:think>" in text or "</mm:think>" in text:
++        text = text.replace("<mm:think>", "<think>").replace("</mm:think>", "</think>")
++
+     thinking_parts = []
+     remaining = text
+ 
+@@ -249,6 +255,11 @@ class ThinkingParser:
+         if not text:
+             return ("", "")
+ 
++        # MiniMax-M3 emits <mm:think>...</mm:think> (each marker is a single token,
++        # so it arrives whole per chunk); normalize to <think>/</think>.
++        if "<mm:think>" in text or "</mm:think>" in text:
++            text = text.replace("<mm:think>", "<think>").replace("</mm:think>", "</think>")
++
+         # Prepend any buffered partial tag content
+         text = self._buffer + text
+         self._buffer = ""

patches/api_tool_calling.py.patch

@@ -1,11 +1,11 @@
---- api/tool_calling.py.bak.minimax-m3-toolcall	2026-06-14 21:39:55
-+++ api/tool_calling.py	2026-06-14 21:43:27
-@@ -464,6 +464,56 @@
-         results.append({"name": func_name, "arguments": arguments})
- 
+diff --git a/api/tool_calling.py b/api/tool_calling.py
+index b51bcbb..9f3f425 100644
+--- a/api/tool_calling.py
++++ b/api/tool_calling.py
+@@ -509,6 +509,56 @@ def _parse_gemma4_tool_call_fallback(text: str) -> Union[dict, list]:
      return results[0] if len(results) == 1 else results
-+
-+
+ 
+ 
 +def _parse_minimax_m3_tool_calls(text):
 +    """Parse MiniMax-M3 (minimax_m3_vl) tool calls.
 +
@@ -54,10 +54,12 @@
 +    cleaned = re.sub(r"<invoke\s+name=.*?</invoke>", "", cleaned, flags=re.DOTALL)
 +    cleaned = cleaned.replace(NS, "").strip()
 +    return cleaned, tool_calls
- 
- 
++
++
  def parse_tool_calls(
-@@ -623,6 +673,12 @@
+     text: str,
+     tokenizer: Any,
+@@ -666,6 +716,12 @@ def parse_tool_calls(
                          cleaned_text = cleaned_text[:idx].strip()
                  return cleaned_text, tool_calls
  

patches/oq.py.patch

@@ -1,6 +1,8 @@
---- oq.py.bak.minimax-m3.20260613	2026-06-13 11:41:21
-+++ oq.py	2026-06-13 11:41:21
-@@ -2030,6 +2030,13 @@
+diff --git a/oq.py b/oq.py
+index 295d9be..06a8268 100644
+--- a/oq.py
++++ b/oq.py
+@@ -1993,6 +1993,13 @@ def _build_model_sanitizer(config: dict, text_only: bool = False):
  
                  proxy = _Proxy()
                  proxy.config = model_config

patches/scheduler.py.patch

@@ -1,8 +1,10 @@
---- scheduler.py.bak.minimax-m3.20260613	2026-06-13 12:41:33
-+++ scheduler.py	2026-06-13 13:15:21
-@@ -577,6 +577,12 @@
-         return type(cache_obj)(*converted)
-     if isinstance(cache_obj, _REGULAR_SINGLETON_CACHE_TYPES):
+diff --git a/scheduler.py b/scheduler.py
+index 2192471..3cb9b12 100644
+--- a/scheduler.py
++++ b/scheduler.py
+@@ -806,6 +806,12 @@ def _to_batched_cache_layer(cache_obj: Any) -> Any:
+         and type(cache_obj) is _TQ_SINGLETON_CACHE_TYPE
+     ):
          return cache_obj.merge([cache_obj])
 +    # MiniMax-M3: convert its single MSA cache to the batch variant
 +    # (MiniMaxM3BatchKVCache implements filter/extend/extract/merge). Without
@@ -13,13 +15,10 @@
      return cache_obj
  
  
-@@ -755,6 +761,54 @@
-             "ChunkedKVCache patch: methods already present upstream, " "skipped: %s",
-             ", ".join(_ckvcache_methods_skipped),
-         )
-+except ImportError:
-+    pass
-+
+@@ -987,6 +993,54 @@ try:
+ except ImportError:
+     pass
+ 
 +# MiniMax-M3 (minimax_m3_vl): its MSA KV cache lacks the batch-aware methods
 +# (filter/extract/merge/extend) oMLX continuous-batching expects, so serving
 +# raises "MiniMaxM3KVCache does not yet support batching with history". M3
@@ -65,10 +64,13 @@
 +                "supported; serialize MiniMax-M3 requests."
 +            )
 +        _M3KV.extend = _m3_extend_passthrough
- except ImportError:
-     pass
++except ImportError:
++    pass
++
  
-@@ -961,7 +1015,7 @@
+ # ---------------------------------------------------------------------------
+ # Monkey-patch PromptProcessingBatch.prompt to set mRoPE deltas before the
+@@ -1207,7 +1261,7 @@ def _model_declares_llama4(model: Any) -> bool:
              continue
          seen.add(obj_id)
  

patches/server.py.patch

@@ -1,6 +1,8 @@
---- server.py.bak.minimax-m3-ns	2026-06-14 21:39:55
-+++ server.py	2026-06-14 21:43:45
-@@ -3969,7 +3969,7 @@
+diff --git a/server.py b/server.py
+index f58b000..88705e2 100644
+--- a/server.py
++++ b/server.py
+@@ -4002,7 +4002,7 @@ async def stream_chat_completion(
                  # tools are present so clients see clean streamed text.
                  if content_delta:
                      if tool_filter:
@@ -9,7 +11,7 @@
                      if content_delta:
                          chunk = ChatCompletionChunk(
                              id=response_id,
-@@ -4031,7 +4031,7 @@
+@@ -4062,7 +4062,7 @@ async def stream_chat_completion(
                  yield f"data: {chunk.model_dump_json(exclude_none=True)}\n\n"
          if content_delta:
              if tool_filter:
@@ -18,7 +20,7 @@
              if content_delta:
                  chunk = ChatCompletionChunk(
                      id=response_id,
-@@ -4046,7 +4046,7 @@
+@@ -4077,7 +4077,7 @@ async def stream_chat_completion(
                  yield f"data: {chunk.model_dump_json(exclude_none=True)}\n\n"
  
          if tool_filter:
@@ -27,7 +29,7 @@
              if remaining:
                  chunk = ChatCompletionChunk(
                      id=response_id,
-@@ -4390,7 +4390,7 @@
+@@ -4419,7 +4419,7 @@ async def stream_anthropic_messages(
                  # markup when a known start marker is available.
                  if content_delta:
                      if tool_filter:
@@ -36,7 +38,7 @@
                      if content_delta:
                          # When tools are requested AND we haven't yet opened
                          # a text block, drop pure-whitespace deltas. Models
-@@ -4467,7 +4467,7 @@
+@@ -4494,7 +4494,7 @@ async def stream_anthropic_messages(
              )
      if content_delta:
          if tool_filter:
@@ -45,7 +47,7 @@
          if content_delta:
              if thinking_block_started and not text_block_started:
                  yield create_content_block_stop_event(index=block_index)
-@@ -4482,7 +4482,7 @@
+@@ -4509,7 +4509,7 @@ async def stream_anthropic_messages(
  
      # Flush any remaining buffered content from the tool-call filter
      if tool_filter:
@@ -54,7 +56,7 @@
          if remaining:
              if not text_block_started:
                  if thinking_block_started:
-@@ -5736,7 +5736,7 @@
+@@ -5774,7 +5774,7 @@ async def stream_responses_api(
                      if msg_output_index is None:
                          msg_output_index = next_output_index
                      if tool_filter:
@@ -63,7 +65,7 @@
                      if content_delta:
                          seq += 1
                          yield format_sse_event(
-@@ -5783,7 +5783,7 @@
+@@ -5819,7 +5819,7 @@ async def stream_responses_api(
              accumulated_reasoning += thinking_delta
          if content_delta:
              if tool_filter:
@@ -72,7 +74,7 @@
              if content_delta:
                  seq += 1
                  yield format_sse_event(
-@@ -5798,7 +5798,7 @@
+@@ -5834,7 +5834,7 @@ async def stream_responses_api(
                      },
                  )
          if tool_filter: