diffsynth-engine 0.4.1.dev1__tar.gz → 0.4.2.dev4__tar.gz

{diffsynth_engine-0.4.1.dev1 → diffsynth_engine-0.4.2.dev4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: diffsynth_engine
-Version: 0.4.1.dev1
+Version: 0.4.2.dev4
 Author: MuseAI x ModelScope
 Classifier: Programming Language :: Python :: 3
 Classifier: Operating System :: OS Independent

{diffsynth_engine-0.4.1.dev1 → diffsynth_engine-0.4.2.dev4}/diffsynth_engine/pipelines/qwen_image.py

@@ -82,10 +82,8 @@ class QwenImagePipeline(BasePipeline):
             dtype=config.model_dtype,
         )
         self.config = config
-        self.tokenizer_max_length = 1024
         self.prompt_template_encode = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n"
         self.prompt_template_encode_start_idx = 34
-        self.default_sample_size = 128
         # sampler
         self.noise_scheduler = RecifitedFlowScheduler(shift=3.0, use_dynamic_shifting=True)
         self.sampler = FlowMatchEulerSampler()
@@ -262,7 +260,7 @@ class QwenImagePipeline(BasePipeline):
         template = self.prompt_template_encode
         drop_idx = self.prompt_template_encode_start_idx
         texts = [template.format(txt) for txt in prompt]
-        outputs = self.tokenizer(texts, max_length=min(max_sequence_length, self.tokenizer_max_length) + drop_idx)
+        outputs = self.tokenizer(texts, max_length=max_sequence_length + drop_idx)
         input_ids, attention_mask = outputs["input_ids"].to(self.device), outputs["attention_mask"].to(self.device)
         outputs = self.encoder(input_ids=input_ids, attention_mask=attention_mask)
         hidden_states = outputs["hidden_states"]
@@ -377,9 +375,9 @@ class QwenImagePipeline(BasePipeline):
         self.sampler.initialize(init_latents=init_latents, timesteps=timesteps, sigmas=sigmas)
 
         self.load_models_to_device(["encoder"])
-        prompt_embeds, prompt_embeds_mask = self.encode_prompt(prompt, 1, 512)
+        prompt_embeds, prompt_embeds_mask = self.encode_prompt(prompt, 1, 4096)
         if cfg_scale > 1.0 and negative_prompt != "":
-            negative_prompt_embeds, negative_prompt_embeds_mask = self.encode_prompt(negative_prompt, 1, 512)
+            negative_prompt_embeds, negative_prompt_embeds_mask = self.encode_prompt(negative_prompt, 1, 4096)
         else:
             negative_prompt_embeds, negative_prompt_embeds_mask = None, None
 

{diffsynth_engine-0.4.1.dev1 → diffsynth_engine-0.4.2.dev4}/diffsynth_engine/tokenizers/qwen2.py

@@ -197,8 +197,8 @@ class Qwen2TokenizerFast(BaseTokenizer):
         encoded.fill_(self.pad_token_id)
         attention_mask = torch.zeros(len(texts), max_length, dtype=torch.long)
         for i, ids in enumerate(batch_ids):
-            if len(ids) > self.model_max_length:
-                ids = ids[: self.model_max_length]
+            if len(ids) > max_length:
+                ids = ids[:max_length]
                 ids[-1] = self.eos_token_id
             if padding_side == "right":
                 encoded[i, : len(ids)] = torch.tensor(ids)

{diffsynth_engine-0.4.1.dev1 → diffsynth_engine-0.4.2.dev4}/diffsynth_engine/utils/parallel.py

@@ -304,12 +304,14 @@ def _worker_loop(
             if rank == 0:
                 queue_out.put(res)
             dist.barrier()
-    except Exception as e:
+    except Exception:
         import traceback
 
-        traceback.print_exc()
-        logger.error(f"Error in worker loop (rank {rank}): {e}")
-        queue_out.put(e)  # any exception caught in the worker will be raised to the main process
+        msg = traceback.format_exc()
+        err = RuntimeError(msg)
+        logger.error(f"Error in worker loop (rank {rank}): {msg}")
+        if rank == 0:
+            queue_out.put(err)  # any exception caught in the worker will be raised to the main process
     finally:
         del module
         torch.cuda.synchronize()

{diffsynth_engine-0.4.1.dev1 → diffsynth_engine-0.4.2.dev4}/diffsynth_engine.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: diffsynth_engine
-Version: 0.4.1.dev1
+Version: 0.4.2.dev4
 Author: MuseAI x ModelScope
 Classifier: Programming Language :: Python :: 3
 Classifier: Operating System :: OS Independent