Please upload - qwen3.8_27b_w4a8_convrot_qhead_mtp.safetensors
Please upload - qwen3.8_27b_w4a8_convrot_qhead_mtp.safetensors
It is this model https://huggingface.co/Comfy-Org/Qwen3.8-27B/blob/main/text_encoders/qwen3.8_27b_w4a8.safetensors
I have been using it but this model doesn't seem to support mtp
It is this model https://huggingface.co/Comfy-Org/Qwen3.8-27B/blob/main/text_encoders/qwen3.8_27b_w4a8.safetensors
I have been using it but this model doesn't seem to support mtp
Double checked and it definitely has the mtp layers.
Kijai is correct. qwen3.8_27b_w4a8.safetensors has MTP support.
By the way, I didn't rely solely on qwen3.8_27b_w4a8.safetensors for this workflow. To generate the precise prompt structure, or rather, the required JSON file I used Llama.cpp running Qwen 3.6. This is actually my go-to local combination now, and I regularly use it to write text and craft highly structured scripts and prompts for video models like LTX 2.5 and MiniMax H3. Having this kind of local pipeline gives you unmatched flexibility. Best of all, I run this entire setup on a completely standard RTX 3060 12GB. And I can get around 50t/sec for Qwen 3.6 35B MoE quant.
Cheers!
It is this model https://huggingface.co/Comfy-Org/Qwen3.8-27B/blob/main/text_encoders/qwen3.8_27b_w4a8.safetensors
I have been using it but this model doesn't seem to support mtp
Double checked and it definitely has the mtp layers.
When I enable MTP on the TextGenerate node with qwen3.8_27b_w4a8.safetensors, it executes up to _generate_mtp(), but crashes during CUDA graph capture:
torch.AcceleratorError: CUDA error: operation failed due to a previous error during capture
(cudaErrorStreamCaptureInvalidated)
Traceback snippet:
File ".../comfy/text_encoders/qwen35.py", line 754, in generate
return self._generate_mtp(embeds, max_length, stop_tokens, sampling=sampling, fixed_depth=fixed_depth)
File ".../comfy/text_encoders/qwen35.py", line 972, in _generate_mtp
draft_capture()
File ".../comfy/text_encoders/qwen35.py", line 870, in draft_capture
with torch.cuda.graph(g, capture_error_mode="thread_local"):
If I disable MTP, generation runs without error, but speed drops dramatically.
Is this weight file missing the dedicated draft head weights required for MTP (which is why I was looking for the _convrot_qhead_mtp build), or is there an issue with draft_capture() CUDA graph recording on newer setups?