query transformers' mapping the way transformers defines it

Reported from production within the hour: `Qwen3.5-9B` now REFUSES pictures,
"served without a vision tower" — the exact opposite of what the vision
loading was for.

transformers' `_LazyAutoMapping.get` is not `dict.get`. Its signature is
`get(self, key, default)` and `default` is REQUIRED, so

    AutoModelForImageTextToText._model_mapping.get(type(cfg))

raises TypeError. `is_vision_checkpoint` caught it and answered "no vision
class" — for every checkpoint, always. Every vision model therefore loaded as
text, and then had its pictures refused by the guard added for exactly that
case. The probe I verified it with passed two arguments, so it tested a call
the code does not make.

Two tests now call what the code calls: one with a stub whose `get` mirrors
the required `default`, and one through the real mapping with a config class
this transformers actually ships.

The silence was the other half. `is_vision_checkpoint` returned False without
a word, so the engine log said nothing at all and the only symptom was a 400
that blamed the model's configuration — which was fine. `vision_report` now
carries the reason at every step, a checkpoint that looks like a vision model
and is not loaded as one says so in the log, and the refusal quotes what was
actually found rather than listing three things it might be.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014T3rZozi1Ehk96NG1WesgC

codai/__init__.py

@@ -16,7 +16,7 @@
 
 # Canonical product version for CoderAI — single source of truth. Both the API
 # metadata and the admin web UI read from here.
-__version__ = "0.5.26"
+__version__ = "0.5.27"
 __tagline__ = "Complete orchestration, distribution and escalation of remotizable advanced inference."
 
 # Configure the CUDA caching allocator BEFORE torch is imported anywhere.

codai/api/text.py

@@ -1638,12 +1638,34 @@ async def chat_completions(request: ChatCompletionRequest, http_request: Request
             if any(isinstance(_it, dict) and _it.get('type') in ('image_url', 'input_image')
                    for _it in _c):
                 _release_instance()
+                # Say what was found, not just what is missing. The first
+                # version of this message named three possible causes and the
+                # engine logged nothing, so when the vision loading itself was
+                # broken (a one-argument call to transformers' mapping, which
+                # raises) the only symptom was this refusal -- and it pointed
+                # at the model's configuration, which was fine.
+                _why = ""
+                try:
+                    from codai.backends import hf_vision as _hv
+                    _r = _hv.vision_report(
+                        getattr(current_manager, 'model_name', None)
+                        or requested_model)
+                    _why = _r.get("reason") or ""
+                    if _r.get("cls") is not None:
+                        _why = ("this checkpoint HAS a vision class "
+                                f"({_r['cls'].__name__ if hasattr(_r['cls'], '__name__') else _r['cls']}) "
+                                "but the loaded model is not using it — the "
+                                "engine log says why")
+                except Exception:
+                    _why = ""
                 raise HTTPException(status_code=400, detail=(
                     f"Model '{requested_model}' cannot read images: it is served "
-                    f"without a vision tower (an HF checkpoint with no "
-                    f"vision_config or processor, or a GGUF with no mmproj). "
-                    f"Send the image to a model whose capabilities list "
-                    f"image_to_text, or configure an mmproj for this one."))
+                    f"without a vision tower"
+                    + (f" ({_why})" if _why else
+                       " (an HF checkpoint with no vision_config or processor, "
+                       "or a GGUF with no mmproj)")
+                    + f". Send the image to a model whose capabilities list "
+                      f"image_to_text, or configure an mmproj for this one."))
 
     # Inject system prompt if --system-prompt flag was provided
     messages = request.messages

codai/backends/hf_vision.py

@@ -73,34 +73,89 @@ def _config_for(model_name: str):
     try:
         from transformers import AutoConfig
         return AutoConfig.from_pretrained(model_name, trust_remote_code=True)
-    except Exception:
+    except Exception as exc:
+        print(f"[vision] {model_name}: its config could not be read "
+              f"({type(exc).__name__}: {str(exc)[:160]})", flush=True)
         return None
 
 
-def is_vision_checkpoint(model_name: str, config=None) -> bool:
-    """Whether this checkpoint carries a vision tower.
+def vision_report(model_name: str, config=None) -> dict:
+    """Why this checkpoint is, or is not, loaded with a vision tower.
 
-    Decided from the config, not from the model id: ``Qwen3.5-9B`` has no
-    "vl"/"vision" in its name and is a vision model, while plenty of models
-    with "vision" in the name are not image-text-to-text at all. The two
-    signals are a nested ``vision_config`` and an ``image_token_id`` -- a
-    decoder that has somewhere to put an image.
+    Every step is reported because every one of them used to fail silently.
+    Measured on zeiss, 2026-10-29: after the vision loading shipped, an image
+    sent to ``Qwen3.5-9B`` was refused with "served without a vision tower" —
+    and the engine logged nothing at all about why, because
+    :func:`is_vision_checkpoint` answered False and said nothing. A decision
+    this consequential has to leave a trace.
+
+    Keys: ``config`` (read or not), ``vision_config``, ``image_token_id``,
+    ``cls`` (the class that would load it, or None), ``reason``.
     """
+    out = {"config": False, "vision_config": False, "image_token_id": False,
+           "cls": None, "reason": ""}
     cfg = config if config is not None else _config_for(model_name)
     if cfg is None:
-        return False
-    if getattr(cfg, "vision_config", None) is None and \
-            getattr(cfg, "image_token_id", None) is None:
-        return False
+        out["reason"] = "its config could not be read"
+        return out
+    out["config"] = True
+    out["cfg_class"] = type(cfg).__name__
+    out["vision_config"] = getattr(cfg, "vision_config", None) is not None
+    out["image_token_id"] = getattr(cfg, "image_token_id", None) is not None
+    if not out["vision_config"] and not out["image_token_id"]:
+        out["reason"] = ("its config has neither a vision_config nor an "
+                         "image_token_id — it is a text checkpoint")
+        return out
     # The config says there is a tower; the installed transformers must also
     # have a class that loads it. Without this check a brand-new family would
     # pick the vision class and fail the load outright, which is worse than
     # serving it as text.
     try:
         from transformers import AutoModelForImageTextToText
-        return AutoModelForImageTextToText._model_mapping.get(type(cfg)) is not None
+        # BOTH arguments. transformers' `_LazyAutoMapping.get` is not dict.get:
+        # its signature is `get(self, key, default)` and `default` is
+        # REQUIRED. Calling it with one argument raises
+        #
+        #   TypeError: _LazyAutoMapping.get() missing 1 required positional
+        #   argument: 'default'
+        #
+        # which this function caught and reported as "no vision class" — for
+        # every checkpoint, always. So every vision model loaded as text and
+        # then had its pictures refused by the guard added for exactly that
+        # case. Measured on zeiss, 2026-10-29, against Qwen3.5-9B.
+        cls = AutoModelForImageTextToText._model_mapping.get(type(cfg), None)
+    except Exception as exc:
+        out["reason"] = (f"the image-text-to-text mapping could not be read "
+                         f"({type(exc).__name__}: {str(exc)[:120]})")
+        return out
+    if cls is None:
+        out["reason"] = (f"this transformers has no image-text-to-text class "
+                         f"for {type(cfg).__name__} ({_tf_version()})")
+        return out
+    out["cls"] = cls
+    return out
+
+
+def _tf_version() -> str:
+    try:
+        import transformers
+        return f"transformers {transformers.__version__}"
     except Exception:
-        return False
+        return "transformers version unknown"
+
+
+def is_vision_checkpoint(model_name: str, config=None) -> bool:
+    """Whether this checkpoint carries a vision tower.
+
+    Decided from the config, not from the model id: ``Qwen3.5-9B`` has no
+    "vl"/"vision" in its name and is a vision model, while plenty of models
+    with "vision" in the name are not image-text-to-text at all. The two
+    signals are a nested ``vision_config`` and an ``image_token_id`` -- a
+    decoder that has somewhere to put an image.
+
+    See :func:`vision_report` for the same decision with its reasons.
+    """
+    return vision_report(model_name, config)["cls"] is not None
 
 
 def auto_model_class(model_name: str):
@@ -112,11 +167,23 @@ def auto_model_class(model_name: str):
     """
     from transformers import AutoModelForCausalLM
     try:
-        if is_vision_checkpoint(model_name):
+        report = vision_report(model_name)
+        if report["cls"] is not None:
             from transformers import AutoModelForImageTextToText
             return AutoModelForImageTextToText, True
-    except Exception:
-        pass
+        # Only worth a line when the checkpoint LOOKS like a vision model and
+        # is not being loaded as one: that is the case an operator has to be
+        # able to see, and the one that used to be silent.
+        if report.get("vision_config") or report.get("image_token_id"):
+            print(f"[vision] {model_name} looks like a vision checkpoint "
+                  f"(vision_config={report['vision_config']}, "
+                  f"image_token_id={report['image_token_id']}, "
+                  f"config={report.get('cfg_class')}) but "
+                  f"{report['reason']} — loading it as text", flush=True)
+    except Exception as exc:
+        print(f"[vision] {model_name}: could not decide on a model class "
+              f"({type(exc).__name__}: {str(exc)[:160]}) — loading it as text",
+              flush=True)
     return AutoModelForCausalLM, False
 
 

tests/test_hf_vision_chat.py

@@ -221,10 +221,15 @@ def test_an_image_to_a_blind_model_is_refused_before_it_is_flattened():
     assert guard < flatten, (
         "the placeholder is built before the refusal, so a blind model can "
         "still answer a picture")
-    body = src[guard:guard + 1400]
+    # The whole guard block, not a fixed slice: a slice breaks the moment the
+    # message grows, which says nothing about the behaviour under test.
+    body = src[guard:src.index("# Inject system prompt", guard)]
     assert "status_code=400" in body
     assert "_release_instance()" in body, (
         "refusing without releasing the instance slot leaks it")
+    assert "vision_report(" in body, (
+        "the refusal does not say what was actually found, which is how a "
+        "broken vision loader looked like a misconfigured model")
 
 
 def test_a_gguf_without_an_mmproj_does_not_advertise_vision():
@@ -234,3 +239,83 @@ def test_a_gguf_without_an_mmproj_does_not_advertise_vision():
     assert '.gguf' in block and 'mmproj' in block
     # a remote/pod-served model is the remote's business, not this listing's
     assert "vllm" in block and "runpod" in block
+
+
+# --------------------------------------------------------------------------
+# The mapping call itself
+# --------------------------------------------------------------------------
+#
+# Shipped broken and found in production the same day: transformers'
+# `_LazyAutoMapping.get` is NOT dict.get — its signature is
+# `get(self, key, default)` with `default` REQUIRED. Called with one argument
+# it raises
+#
+#   TypeError: _LazyAutoMapping.get() missing 1 required positional argument
+#
+# which this module caught and reported as "no vision class" — for every
+# checkpoint, always. So every vision model loaded as text and then had its
+# pictures refused by the guard added for exactly that case.
+#
+# The probe that "verified" it passed two arguments, so it tested a call the
+# code does not make. These tests call what the code calls.
+
+
+def test_the_mapping_is_queried_the_way_transformers_defines_it():
+    """A stub whose `get` mirrors _LazyAutoMapping's required `default`."""
+    import transformers
+    from codai.backends import hf_vision
+
+    class _Cfg:
+        vision_config = {"depth": 2}
+        image_token_id = 7
+
+    class _Cls:
+        pass
+
+    class _StrictMapping:
+        def get(self, key, default):          # default is REQUIRED, as upstream
+            return {_Cfg: _Cls}.get(key, default)
+
+    class _Auto:
+        _model_mapping = _StrictMapping()
+
+    real = transformers.AutoModelForImageTextToText
+    try:
+        transformers.AutoModelForImageTextToText = _Auto
+        report = hf_vision.vision_report("x/y", config=_Cfg())
+    finally:
+        transformers.AutoModelForImageTextToText = real
+    assert report["cls"] is _Cls, report["reason"]
+
+
+def test_a_real_installed_vision_config_resolves_to_a_class():
+    """No stub, no network, no weights: a config class this transformers
+    actually ships, through the real mapping."""
+    import transformers
+    from codai.backends import hf_vision
+
+    cfg = None
+    for name in ("LlavaConfig", "Qwen2VLConfig", "Blip2Config", "PaliGemmaConfig"):
+        klass = getattr(transformers, name, None)
+        if klass is None:
+            continue
+        try:
+            cfg = klass()
+        except Exception:
+            continue
+        break
+    if cfg is None:
+        pytest.skip("this transformers ships none of the probe configs")
+    report = hf_vision.vision_report("x/y", config=cfg)
+    assert report["cls"] is not None, report["reason"]
+
+
+def test_a_failure_to_decide_is_never_silent():
+    """The 400 that refuses a picture was the only trace of this. A decision
+    that consequential has to leave one of its own."""
+    src = _src("codai/backends/hf_vision.py")
+    rep = src.split("def vision_report", 1)[1].split("\ndef ", 1)[0]
+    assert '"reason"' in rep
+    auto = src.split("def auto_model_class", 1)[1].split("\ndef ", 1)[0]
+    assert "print(" in auto, (
+        "a vision checkpoint loaded as text says nothing in the log")