From fef9505848e3385622a60cf64c7f081a02f96b5a Mon Sep 17 00:00:00 2001
From: Oleksandr Pichak <stasselust@gmail.com>
Date: Mon, 24 Aug 2026 12:32:15 +0200
Subject: [PATCH] fix(classifier): adopt the tokenizer pad_token_id when the
 model config lacks one

Decoder-only sequence-classification heads (Qwen2/Qwen3ForSequenceClassification)
raise `Cannot handle batch sizes > 1 if no padding token is defined.` when
config.pad_token_id is unset. michaelfeil/Qwen3-Reranker-0.6B-seq, produced by
docs/lm_head_to_classifier/convert_lm.py, ships without that key, so infinity
crashes during warmup at batch_size=32 before serving a request.

No-op for every checkpoint that already carries pad_token_id.
---
 .../transformer/classifier/torch.py           | 20 +++++++++
 .../classifier/test_torch_classifer.py        | 41 ++++++++++++++++++-
 2 files changed, 60 insertions(+), 1 deletion(-)

diff --git a/libs/infinity_emb/infinity_emb/transformer/classifier/torch.py b/libs/infinity_emb/infinity_emb/transformer/classifier/torch.py
index 3c9e7045e..97f21fdb1 100644
--- a/libs/infinity_emb/infinity_emb/transformer/classifier/torch.py
+++ b/libs/infinity_emb/infinity_emb/transformer/classifier/torch.py
@@ -18,6 +18,24 @@
     import torch
 
 
+def _set_pad_token_id_if_missing(model, tokenizer) -> None:
+    """Copy the tokenizer's pad_token_id onto the model config when it is missing.
+
+    Decoder-only sequence-classification heads (e.g. Qwen2/Qwen3ForSequenceClassification)
+    raise `Cannot handle batch sizes > 1 if no padding token is defined.` when
+    `config.pad_token_id` is unset, and several published checkpoints ship without it.
+    """
+    if getattr(model.config, "pad_token_id", None) is not None:
+        return
+    pad_token_id = getattr(tokenizer, "pad_token_id", None)
+    if pad_token_id is None:
+        return
+    logger.info(
+        f"model config has no pad_token_id, adopting the tokenizer's pad_token_id={pad_token_id}"
+    )
+    model.config.pad_token_id = pad_token_id
+
+
 class SentenceClassifier(BaseClassifer):
     def __init__(
         self,
@@ -45,6 +63,8 @@ def __init__(
             model_kwargs=model_kwargs,
         )
 
+        _set_pad_token_id_if_missing(self._pipe.model, self._pipe.tokenizer)
+
         if ls.quantization_dtype is not None:
             self._pipe.model = quant_interface(  # TODO: add ls.quantization_dtype and ls.placement
                 self._pipe.model, engine_args.dtype, device=Device[self._pipe.model.device.type]
diff --git a/libs/infinity_emb/tests/unit_test/transformer/classifier/test_torch_classifer.py b/libs/infinity_emb/tests/unit_test/transformer/classifier/test_torch_classifer.py
index 814536d74..653e5ed3c 100644
--- a/libs/infinity_emb/tests/unit_test/transformer/classifier/test_torch_classifer.py
+++ b/libs/infinity_emb/tests/unit_test/transformer/classifier/test_torch_classifer.py
@@ -1,8 +1,13 @@
+from types import SimpleNamespace
+
 import torch
 from transformers import pipeline  # type: ignore
 
 from infinity_emb.args import EngineArgs
-from infinity_emb.transformer.classifier.torch import SentenceClassifier
+from infinity_emb.transformer.classifier.torch import (
+    SentenceClassifier,
+    _set_pad_token_id_if_missing,
+)
 
 
 def test_classifier(model_name: str = "SamLowe/roberta-base-go_emotions"):
@@ -40,3 +45,37 @@ def test_classifier(model_name: str = "SamLowe/roberta-base-go_emotions"):
 
             if pred_orig_i["score"] > 0.005:
                 assert pred_orig_i["label"] == pred_i["label"]
+
+
+def test_set_pad_token_id_if_missing_adopts_tokenizer_value():
+    """decoder-only seq-cls checkpoints often ship without config.pad_token_id.
+
+    transformers then raises `Cannot handle batch sizes > 1 if no padding token is
+    defined.`, which infinity hits during warmup (batch_size 32) before serving a
+    single request. michaelfeil/Qwen3-Reranker-0.6B-seq is such a checkpoint.
+    """
+    model = SimpleNamespace(config=SimpleNamespace(pad_token_id=None))
+    tokenizer = SimpleNamespace(pad_token_id=151643)
+
+    _set_pad_token_id_if_missing(model, tokenizer)
+
+    assert model.config.pad_token_id == 151643
+
+
+def test_set_pad_token_id_if_missing_is_noop_when_already_set():
+    """strict no-op for every model that works today (BERT rerankers, mxbai-*-seq)."""
+    model = SimpleNamespace(config=SimpleNamespace(pad_token_id=0))
+    tokenizer = SimpleNamespace(pad_token_id=151643)
+
+    _set_pad_token_id_if_missing(model, tokenizer)
+
+    assert model.config.pad_token_id == 0
+
+
+def test_set_pad_token_id_if_missing_tolerates_tokenizer_without_pad():
+    model = SimpleNamespace(config=SimpleNamespace(pad_token_id=None))
+    tokenizer = SimpleNamespace(pad_token_id=None)
+
+    _set_pad_token_id_if_missing(model, tokenizer)
+
+    assert model.config.pad_token_id is None