Skip to content

The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.

Fix Qwen think-block templates in TextGenerate (#17121)

Jukka Seppänence1e922d8623
diff --git a/comfy/sd.py b/comfy/sd.py
index 6d5ef40c5..e0b9d276b 100644
--- a/comfy/sd.py
+++ b/comfy/sd.py
@@ -1965,7 +1965,7 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
elif te_model in (TEModel.QWEN35_08B, TEModel.QWEN35_2B, TEModel.QWEN35_4B, TEModel.QWEN35_9B, TEModel.QWEN35_27B):
clip_data[0] = comfy.utils.state_dict_prefix_replace(clip_data[0], {"model.language_model.": "model.", "model.visual.": "visual.", "lm_head.": "model.lm_head."})
if clip_type == CLIPType.MOSS_IMAGE and te_model == TEModel.QWEN35_9B: # MOSS-Image: stock Qwen3.5-9B, last hidden state, open think block.
- clip_target.clip = comfy.text_encoders.moss_image.te(**llama_detect(clip_data))
+ clip_target.clip = comfy.text_encoders.moss_image.te(**llama_detect(clip_data), mtp="mtp.fc.weight" in clip_data[0])
clip_target.tokenizer = comfy.text_encoders.moss_image.MOSSImageTokenizer
else:
qwen35_type = {TEModel.QWEN35_08B: "qwen35_08b", TEModel.QWEN35_2B: "qwen35_2b", TEModel.QWEN35_4B: "qwen35_4b", TEModel.QWEN35_9B: "qwen35_9b", TEModel.QWEN35_27B: "qwen35_27b"}[te_model]
diff --git a/comfy/text_encoders/boogu.py b/comfy/text_encoders/boogu.py
index d9de92f10..65eb61545 100644
--- a/comfy/text_encoders/boogu.py
+++ b/comfy/text_encoders/boogu.py
@@ -26,11 +26,10 @@ class BooguTokenizer(comfy.text_encoders.qwen3vl.Qwen3VLTokenizer):
# Reference SYSTEM_PROMPT_DROP: used for the empty negative/uncond instruction.
self.llama_template_drop = "<|im_start|>system\n" + BOOGU_DROP_SYSTEM + "<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n"
- def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, images=[], prevent_empty_text=False, thinking=True, **kwargs):
+ def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, images=[], prevent_empty_text=False, **kwargs):
if llama_template is None and len(images) == 0 and text.strip() == "":
llama_template = self.llama_template_drop
- # Boogu conditions on the no-think template; thinking=True drops the empty <think> block qwen3vl adds by default.
- return super().tokenize_with_weights(text, return_word_ids=return_word_ids, llama_template=llama_template, images=images, prevent_empty_text=prevent_empty_text, thinking=thinking, **kwargs)
+ return super().tokenize_with_weights(text, return_word_ids=return_word_ids, llama_template=llama_template, images=images, prevent_empty_text=prevent_empty_text, **kwargs)
class BooguQwen3VLClipModel(comfy.text_encoders.qwen3vl.Qwen3VLClipModel):
diff --git a/comfy/text_encoders/flux.py b/comfy/text_encoders/flux.py
index 26a596dba..5658740d0 100644
--- a/comfy/text_encoders/flux.py
+++ b/comfy/text_encoders/flux.py
@@ -161,9 +161,11 @@ class KleinTokenizer(sd1_clip.SD1Tokenizer):
super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name=name, tokenizer=tokenizer)
self.llama_template = "<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n"
- def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, system_prompt="", **kwargs):
+ def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, system_prompt="", thinking=False, **kwargs):
if llama_template is None:
llama_text = self.llama_template.format(text)
+ if thinking: # hybrid Qwen3 chat template: the model opens its own think block
+ llama_text = llama_text.removesuffix("<think>\n\n</think>\n\n")
else:
llama_text = llama_template.format(text)
if system_prompt:
@@ -204,9 +206,9 @@ class Flux3Tokenizer(comfy.text_encoders.qwen3vl.Qwen3VLTokenizer):
def __init__(self, embedding_directory=None, tokenizer_data={}):
super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, model_type="qwen3vl_4b")
- def tokenize_with_weights(self, text, return_word_ids=False, thinking=True, **kwargs):
- # instruct template without a think block, right padded to a multiple of 80 tokens
- tokens = super().tokenize_with_weights(text, return_word_ids=return_word_ids, thinking=thinking, **kwargs)
+ def tokenize_with_weights(self, text, return_word_ids=False, **kwargs):
+ # instruct template, right padded to a multiple of 80 tokens
+ tokens = super().tokenize_with_weights(text, return_word_ids=return_word_ids, **kwargs)
pad = (151643, 1.0, 0) if return_word_ids else (151643, 1.0)
for batch in tokens[self.clip_name]:
batch.extend([pad] * (-len(batch) % 80))
diff --git a/comfy/text_encoders/ideogram4.py b/comfy/text_encoders/ideogram4.py
index 151b43c53..cc23e2391 100644
--- a/comfy/text_encoders/ideogram4.py
+++ b/comfy/text_encoders/ideogram4.py
@@ -103,10 +103,6 @@ class Ideogram4Qwen3VLTokenizer(comfy.text_encoders.qwen3vl.Qwen3VLTokenizer):
def __init__(self, embedding_directory=None, tokenizer_data={}):
super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, model_type="qwen3vl_8b")
- def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, images=[], prevent_empty_text=False, thinking=True, **kwargs):
- # Ideogram 4 conditions on the no-think template; default thinking=True drops the empty think block qwen3vl adds.
- return super().tokenize_with_weights(text, return_word_ids=return_word_ids, llama_template=llama_template, images=images, prevent_empty_text=prevent_empty_text, thinking=thinking, **kwargs)
-
def te_qwen3vl(dtype_llama=None, llama_quantization_metadata=None):
class Ideogram4Qwen3VLTEModel_(Ideogram4Qwen3VLTEModel):
diff --git a/comfy/text_encoders/joyimage.py b/comfy/text_encoders/joyimage.py
index 143c44250..5d22c3599 100644
--- a/comfy/text_encoders/joyimage.py
+++ b/comfy/text_encoders/joyimage.py
@@ -45,10 +45,9 @@ class JoyImageTokenizer(Qwen3VLTokenizer):
self.llama_template_images = JOYIMAGE_TEMPLATE_IMAGE
def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, images=None, **kwargs):
- kwargs.pop("thinking", None)
return super().tokenize_with_weights(
text, return_word_ids=return_word_ids, llama_template=llama_template,
- images=images or [], thinking=True, **kwargs,
+ images=images or [], **kwargs,
)
diff --git a/comfy/text_encoders/krea2.py b/comfy/text_encoders/krea2.py
index 408a03566..fc999ad7a 100644
--- a/comfy/text_encoders/krea2.py
+++ b/comfy/text_encoders/krea2.py
@@ -25,10 +25,6 @@ class Krea2Tokenizer(comfy.text_encoders.qwen3vl.Qwen3VLTokenizer):
super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, model_type="qwen3vl_4b")
self.llama_template = KREA2_TEMPLATE # conditioning template; image text-gen uses qwen3vl's default image template.
- def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, images=[], prevent_empty_text=False, thinking=True, **kwargs):
- # Krea2 conditions on the no-think template; thinking=True drops the empty <think> block qwen3vl adds.
- return super().tokenize_with_weights(text, return_word_ids=return_word_ids, llama_template=llama_template, images=images, prevent_empty_text=prevent_empty_text, thinking=thinking, **kwargs)
-
class Krea2Qwen3VLClipModel(comfy.text_encoders.qwen3vl.Qwen3VLClipModel):
def __init__(self, device="cpu", dtype=None, attention_mask=True, model_options={}):
diff --git a/comfy/text_encoders/mage_flow.py b/comfy/text_encoders/mage_flow.py
index 6542ad315..9813fadaa 100644
--- a/comfy/text_encoders/mage_flow.py
+++ b/comfy/text_encoders/mage_flow.py
@@ -26,7 +26,7 @@ class MageFlowTokenizer(comfy.text_encoders.qwen3vl.Qwen3VLTokenizer):
self.llama_template = MAGE_T2I_TEMPLATE
self.llama_template_images = MAGE_EDIT_TEMPLATE
- def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, images=[], prevent_empty_text=False, thinking=True, **kwargs):
+ def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, images=[], prevent_empty_text=False, **kwargs):
image = kwargs.get("image", None)
if image is not None and len(images) == 0:
images = [image[i:i + 1] for i in range(image.shape[0])]
@@ -37,8 +37,7 @@ class MageFlowTokenizer(comfy.text_encoders.qwen3vl.Qwen3VLTokenizer):
llama_template = self.llama_template_images.replace("{}", prefix + "{}", 1)
else:
llama_template = self.llama_template
- # thinking=True: Mage templates end at "<|im_start|>assistant\n" with no <think> block.
- return super().tokenize_with_weights(text, return_word_ids=return_word_ids, llama_template=llama_template, images=images, prevent_empty_text=prevent_empty_text, thinking=thinking, **kwargs)
+ return super().tokenize_with_weights(text, return_word_ids=return_word_ids, llama_template=llama_template, images=images, prevent_empty_text=prevent_empty_text, **kwargs)
class MageFlowQwen3VLClipModel(comfy.text_encoders.qwen3vl.Qwen3VLClipModel):
diff --git a/comfy/text_encoders/moss_image.py b/comfy/text_encoders/moss_image.py
index bd26280ba..070e084ee 100644
--- a/comfy/text_encoders/moss_image.py
+++ b/comfy/text_encoders/moss_image.py
@@ -14,12 +14,12 @@ class MOSSImageTokenizer(comfy.text_encoders.qwen35.Qwen35ImageTokenizer):
class MOSSImageTEModel(sd1_clip.SD1ClipModel):
- def __init__(self, device="cpu", dtype=None, model_options={}):
- clip_model = lambda **kw: comfy.text_encoders.qwen35.Qwen35ClipModel(**kw, layer="last", layer_idx=None, model_type="qwen35_9b")
+ def __init__(self, device="cpu", dtype=None, model_options={}, mtp=False):
+ clip_model = lambda **kw: comfy.text_encoders.qwen35.Qwen35ClipModel(**kw, layer="last", layer_idx=None, model_type="qwen35_9b", mtp=mtp)
super().__init__(device=device, dtype=dtype, name="qwen35_9b", clip_model=clip_model, model_options=model_options)
-def te(dtype_llama=None, llama_quantization_metadata=None):
+def te(dtype_llama=None, llama_quantization_metadata=None, mtp=False):
class MOSSImageTEModel_(MOSSImageTEModel):
def __init__(self, device="cpu", dtype=None, model_options={}):
if dtype_llama is not None:
@@ -27,5 +27,5 @@ def te(dtype_llama=None, llama_quantization_metadata=None):
if llama_quantization_metadata is not None:
model_options = model_options.copy()
model_options["quantization_metadata"] = llama_quantization_metadata
- super().__init__(device=device, dtype=dtype, model_options=model_options)
+ super().__init__(device=device, dtype=dtype, model_options=model_options, mtp=mtp)
return MOSSImageTEModel_
diff --git a/comfy/text_encoders/qwen35.py b/comfy/text_encoders/qwen35.py
index c114869bb..c1e7d08e6 100644
--- a/comfy/text_encoders/qwen35.py
+++ b/comfy/text_encoders/qwen35.py
@@ -1062,8 +1062,8 @@ class Qwen35ImageTokenizer(sd1_clip.SD1Tokenizer):
llama_text = template.format(text)
if system_prompt:
llama_text = "<|im_start|>system\n" + system_prompt + "<|im_end|>\n" + llama_text
- if not thinking:
- llama_text += "<think>\n</think>\n"
+ if llama_template is None: # chat template: thinking opens the block, otherwise an empty block disables reasoning
+ llama_text += "<think>\n" if thinking else "<think>\n\n</think>\n\n"
tokens = super().tokenize_with_weights(llama_text, return_word_ids=return_word_ids, disable_weights=True, **kwargs)
key_name = next(iter(tokens))
diff --git a/comfy/text_encoders/qwen3vl.py b/comfy/text_encoders/qwen3vl.py
index 050e76270..d1d869a26 100644
--- a/comfy/text_encoders/qwen3vl.py
+++ b/comfy/text_encoders/qwen3vl.py
@@ -159,7 +159,7 @@ class Qwen3VLTokenizer(sd1_clip.SD1Tokenizer):
self.llama_template = "<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n"
self.llama_template_images = "<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>{}<|im_end|>\n<|im_start|>assistant\n"
- def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, images=[], prevent_empty_text=False, thinking=False, skip_template=False, system_prompt="", **kwargs):
+ def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, images=[], prevent_empty_text=False, skip_template=False, system_prompt="", **kwargs):
image = kwargs.get("image", None)
if image is not None and len(images) == 0:
images = [image[i:i + 1] for i in range(image.shape[0])]
@@ -183,8 +183,6 @@ class Qwen3VLTokenizer(sd1_clip.SD1Tokenizer):
llama_text = template.format(text)
if system_prompt: # replaces any system turn the template carries
llama_text = "<|im_start|>system\n" + system_prompt + "<|im_end|>\n" + llama_text[llama_text.index("<|im_start|>user"):]
- if not thinking: # Qwen3 convention: empty think block suppresses reasoning
- llama_text += "<think>\n\n</think>\n\n"
tokens = super().tokenize_with_weights(llama_text, return_word_ids=return_word_ids, disable_weights=True, **kwargs)
key_name = next(iter(tokens))
diff --git a/comfy/text_encoders/qwen_image21.py b/comfy/text_encoders/qwen_image21.py
index de0d147d4..25fb221ca 100644
--- a/comfy/text_encoders/qwen_image21.py
+++ b/comfy/text_encoders/qwen_image21.py
@@ -15,14 +15,14 @@ class QwenImage21Tokenizer(comfy.text_encoders.qwen3vl.Qwen3VLTokenizer):
super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, model_type="qwen3vl_8b")
self.llama_template = T2I_TEMPLATE
- def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, images=[], prevent_empty_text=False, thinking=True, keep_vision=False, **kwargs):
+ def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, images=[], prevent_empty_text=False, keep_vision=False, **kwargs):
image = kwargs.get("image", None)
if image is not None and len(images) == 0:
images = [image[i:i + 1] for i in range(image.shape[0])]
if llama_template is None and len(images) > 0:
refs = " ".join("<image{}>{}".format(i + 1, VISION_BLOCK) for i in range(len(images)))
llama_template = T2I_TEMPLATE.replace("{}", refs + "{}", 1)
- out = super().tokenize_with_weights(text, return_word_ids=return_word_ids, llama_template=llama_template, images=images, prevent_empty_text=prevent_empty_text, thinking=thinking, **kwargs)
+ out = super().tokenize_with_weights(text, return_word_ids=return_word_ids, llama_template=llama_template, images=images, prevent_empty_text=prevent_empty_text, **kwargs)
out["keep_vision"] = keep_vision
return out
diff --git a/comfy/text_encoders/z_image.py b/comfy/text_encoders/z_image.py
index 498882d6e..953469f4a 100644
--- a/comfy/text_encoders/z_image.py
+++ b/comfy/text_encoders/z_image.py
@@ -14,9 +14,11 @@ class ZImageTokenizer(sd1_clip.SD1Tokenizer):
super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="qwen3_4b", tokenizer=Qwen3Tokenizer)
self.llama_template = "<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n"
- def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, system_prompt="", **kwargs):
+ def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, system_prompt="", thinking=True, **kwargs):
if llama_template is None:
llama_text = self.llama_template.format(text)
+ if not thinking: # hybrid Qwen3 chat template: conditioning uses the thinking form, an empty block disables reasoning
+ llama_text += "<think>\n\n</think>\n\n"
else:
llama_text = llama_template.format(text)
if system_prompt:
diff --git a/comfy_extras/nodes_textgen.py b/comfy_extras/nodes_textgen.py
index ae1ead89b..efd42b98b 100644
--- a/comfy_extras/nodes_textgen.py
+++ b/comfy_extras/nodes_textgen.py
@@ -81,10 +81,14 @@ class TextGenerate(io.ComfyNode):
generated_text = clip.decode(generated_ids)
- # Reasoning opens in the generated text or at the end of a raw prompt, and has no close when max_length cut it short.
- reasoning, _, text = generated_text.partition("</think>")
- if not reasoning.lstrip().startswith("<think>") and not prompt.rstrip().endswith("<think>"):
- reasoning, text = "", generated_text
+ # </think> ends reasoning; without one, reasoning opened by the output or prompt was cut off by max_length
+ reasoning, sep, text = generated_text.partition("</think>")
+ if not sep and not reasoning.lstrip().startswith("<think>"):
+ if isinstance(tokens, dict):
+ tokens = next(iter(tokens.values()))
+ prompt_ids = [t[0] for t in tokens[-1] if isinstance(t, tuple) and isinstance(t[0], int)]
+ if not prompt.rstrip().endswith("<think>") and not clip.decode(prompt_ids[-4:]).rstrip().endswith("<think>"):
+ reasoning, text = "", generated_text
return io.NodeOutput(text.strip(), reasoning.replace("<think>", "", 1).strip())