| |
| |
| |
| |
| @@ -1,7 +1,10 @@ |
| +import json |
| import numbers |
| |
| import torch |
| |
| +import comfy.ops |
| +import comfy.model_management |
| import comfy.text_encoders.qwen3vl |
| from comfy import sd1_clip |
| |
| @@ -33,9 +36,25 @@ class QwenImage21Qwen3VLClipModel(comfy.text_encoders.qwen3vl.Qwen3VLClipModel): |
| # last layer without the final RMSNorm: transformers 4.57 hidden_states[-1], which Qwen's results are tuned to (5.x norms it) |
| self.layer_norm_hidden_state = False |
| self.image_spans = [] |
| + self.nvfp4_conditioning = False |
| + |
| + def load_sd(self, sd): |
| + self.nvfp4_conditioning = any(json.loads(v.numpy().tobytes()).get("format") == "nvfp4" for k, v in sd.items() if k.endswith(".comfy_quant")) |
| + return super().load_sd(sd) |
| + |
| + def forward(self, tokens): |
| + device = self.execution_device |
| + if device is None: |
| + device = self.transformer.get_input_embeddings().weight.device |
| + if not self.nvfp4_conditioning or not comfy.model_management.supports_nvfp4_compute(device): |
| + return super().forward(tokens) |
| + with comfy.ops.use_quantized_matmul(self.transformer, device): |
| + return super().forward(tokens) |
| |
| def process_tokens(self, tokens, device): |
| embeds, attention_mask, num_tokens, embeds_info = super().process_tokens(tokens, device) |
| + if self.nvfp4_conditioning and comfy.model_management.supports_nvfp4_compute(device): |
| + embeds = embeds.to(dtype=torch.bfloat16) |
| self.image_spans = [(e["index"], e["size"]) for e in embeds_info if e["type"] == "image"] |
| return embeds, attention_mask, num_tokens, embeds_info |
| |
|
|
