diffsynth-engine 0.6.1.dev33__py3-none-any.whl → 0.6.1.dev34__py3-none-any.whl

diffsynth_engine/models/basic/attention.py

@@ -343,7 +343,7 @@ def long_context_attention(
                     f"head_dim={q.shape[-1]}, but aiter_flash_attn only supports head dimension at most {FA3_MAX_HEADDIM}, will use fallback attention implementation"
                 )
         if SDPA_AVAILABLE:
-            return LongContextAttention(attn_type=AttnType.TORCH)(q, k, v, softmax_scale=scale)
+            return LongContextAttention(attn_type=AttnType.TORCH_EFFICIENT)(q, k, v, softmax_scale=scale)
         if FLASH_ATTN_2_AVAILABLE:
             return LongContextAttention(attn_type=AttnType.FA)(q, k, v, softmax_scale=scale)
         raise ValueError("No available long context attention implementation")
@@ -379,7 +379,7 @@ def long_context_attention(
         if attn_impl == "fa2":
             return LongContextAttention(attn_type=AttnType.FA)(q, k, v, softmax_scale=scale)
         if attn_impl == "sdpa":
-            return LongContextAttention(attn_type=AttnType.TORCH)(q, k, v, softmax_scale=scale)
+            return LongContextAttention(attn_type=AttnType.TORCH_EFFICIENT)(q, k, v, softmax_scale=scale)
         if attn_impl == "sage":
             return LongContextAttention(attn_type=AttnType.SAGE_AUTO)(q, k, v, softmax_scale=scale)
         if attn_impl == "sparge":

diffsynth_engine/models/qwen_image/qwen_image_dit.py

@@ -286,16 +286,15 @@ class QwenImageTransformerBlock(nn.Module):
             shift_0, shift_1 = shift[:actual_batch], shift[actual_batch:]
             scale_0, scale_1 = scale[:actual_batch], scale[actual_batch:]
             gate_0, gate_1 = gate[:actual_batch], gate[actual_batch:]
-            index_expanded = index.unsqueeze(-1)
             shift_0_exp = shift_0.unsqueeze(1)
             shift_1_exp = shift_1.unsqueeze(1)
             scale_0_exp = scale_0.unsqueeze(1)
             scale_1_exp = scale_1.unsqueeze(1)
             gate_0_exp = gate_0.unsqueeze(1)
             gate_1_exp = gate_1.unsqueeze(1)
-            shift_result = torch.where(index_expanded == 0, shift_0_exp, shift_1_exp)
-            scale_result = torch.where(index_expanded == 0, scale_0_exp, scale_1_exp)
-            gate_result = torch.where(index_expanded == 0, gate_0_exp, gate_1_exp)
+            shift_result = torch.where(index == 0, shift_0_exp, shift_1_exp)
+            scale_result = torch.where(index == 0, scale_0_exp, scale_1_exp)
+            gate_result = torch.where(index == 0, gate_0_exp, gate_1_exp)
         else:
             shift_result = shift.unsqueeze(1)
             scale_result = scale.unsqueeze(1)
@@ -514,6 +513,7 @@ class QwenImageDiT(PreTrainedModel):
                     device=timestep.device,
                     dtype=torch.int,
                 )
+                modulate_index = modulate_index.unsqueeze(-1)
             rotary_emb = self.pos_embed(video_fhw, text_seq_len, image.device)
 
             image = self.img_in(image)
@@ -535,7 +535,7 @@ class QwenImageDiT(PreTrainedModel):
 
             # warning: Eligen does not work with sequence parallel because long context attention does not support attention masks
             img_freqs, txt_freqs = rotary_emb
-            with sequence_parallel((image, text, img_freqs, txt_freqs), seq_dims=(1, 1, 0, 0)):
+            with sequence_parallel((image, text, img_freqs, txt_freqs, modulate_index), seq_dims=(1, 1, 0, 0, 1)):
                 rotary_emb = (img_freqs, txt_freqs)
                 for block in self.transformer_blocks:
                     text, image = block(

diffsynth_engine/models/qwen_image/qwen_image_vae.py

@@ -685,7 +685,6 @@ class VideoVAE(nn.Module):
         x = patchify(x, patch_size=2 if self.in_channels == 12 else 1)
         t = x.shape[2]
         iter_ = 1 + (t - 1) // 4
-
         for i in range(iter_):
             if i == 0:
                 out = self.encoder(x[:, :, :1, :, :], feat_cache=feat_cache)

diffsynth_engine/pipelines/qwen_image.py

@@ -165,7 +165,7 @@ class QwenImagePipeline(BasePipeline):
         self.edit_prompt_template_encode_start_idx = 64
 
         # sampler
-        self.noise_scheduler = RecifitedFlowScheduler(shift=3.0, use_dynamic_shifting=True)
+        self.noise_scheduler = RecifitedFlowScheduler(shift=3.0, use_dynamic_shifting=True, shift_terminal=0.02)
         self.sampler = FlowMatchEulerSampler()
         # models
         self.tokenizer = tokenizer
@@ -690,8 +690,9 @@ class QwenImagePipeline(BasePipeline):
                 img_width, img_height = img.size
                 condition_width, condition_height = self.calculate_dimensions(384 * 384, img_width / img_height)
                 vae_width, vae_height = self.calculate_dimensions(1024 * 1024, img_width / img_height)
-                condition_images.append(img.resize((condition_width, condition_height), Image.LANCZOS))
-                vae_images.append(img.resize((vae_width, vae_height), Image.LANCZOS))
+                condition_images.append(img.resize((condition_width, condition_height)))
+                vae_images.append(img.resize((vae_width, vae_height)))
+
             if width is None and height is None:
                 width, height = vae_images[-1].size
 

{diffsynth_engine-0.6.1.dev33.dist-info → diffsynth_engine-0.6.1.dev34.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: diffsynth_engine
-Version: 0.6.1.dev33
+Version: 0.6.1.dev34
 Author: MuseAI x ModelScope
 Classifier: Programming Language :: Python :: 3
 Classifier: Operating System :: OS Independent

{diffsynth_engine-0.6.1.dev33.dist-info → diffsynth_engine-0.6.1.dev34.dist-info}/RECORD

@@ -86,7 +86,7 @@ diffsynth_engine/kernels/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZ
 diffsynth_engine/models/__init__.py,sha256=8Ze7cSE8InetgXWTNb0neVA2Q44K7WlE-h7O-02m2sY,119
 diffsynth_engine/models/base.py,sha256=svao__9WH8VNcyXz5o5dzywYXDcGV0YV9IfkLzDKews,2558
 diffsynth_engine/models/basic/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-diffsynth_engine/models/basic/attention.py,sha256=62Ar8_ydnn28F1qH9ueXtvISgNszQK3q8k14gCIXGEs,15681
+diffsynth_engine/models/basic/attention.py,sha256=YrIxkYoekC3I7-sMTw60CL4GIKMLOTrn-eCk-iHT7E4,15701
 diffsynth_engine/models/basic/lora.py,sha256=Y6cBgrBsuDAP9FZz_fgK8vBi_EMg23saFIUSAsPIG-M,10670
 diffsynth_engine/models/basic/lora_nunchaku.py,sha256=7qhzGCzUIfDrwtWG0nspwdyZ7YUkaM4vMqzxZby2Zds,7510
 diffsynth_engine/models/basic/relative_position_emb.py,sha256=rCXOweZMcayVnNUVvBcYXMdhHS257B_PC8PZSWxvhNQ,2540
@@ -111,10 +111,10 @@ diffsynth_engine/models/hunyuan3d/surface_extractor.py,sha256=b15mb1N4PYwAvDk1Gu
 diffsynth_engine/models/hunyuan3d/volume_decoder.py,sha256=sgflj1a8sIerqGSalBAVQOlyiIihkLOLXYysNbulCoQ,2355
 diffsynth_engine/models/qwen_image/__init__.py,sha256=_6f0LWaoLdDvD2CsjK2OzEIQryt9efge8DFS4_GUnHQ,582
 diffsynth_engine/models/qwen_image/qwen2_5_vl.py,sha256=Eu-r-c42t_q74Qpwz21ToCGHpvSi7VND4B1EI0e-ePA,57748
-diffsynth_engine/models/qwen_image/qwen_image_dit.py,sha256=JEyK_yOa0A5xaqlmxI3nfD7NdCaHuvLDA10aWVbnac4,24635
+diffsynth_engine/models/qwen_image/qwen_image_dit.py,sha256=mMU4zeZi8-uJe9voznNIxZCTCqJPbPXkMxHwgcqJ6z8,24640
 diffsynth_engine/models/qwen_image/qwen_image_dit_fbcache.py,sha256=LIv9X_BohKk5rcEzyl3ATLwd8MSoFX43wjkArQ68nq8,4828
 diffsynth_engine/models/qwen_image/qwen_image_dit_nunchaku.py,sha256=1y1BkPRrX4_RioKjM09D9f9PK9neug1nSGJka0D9bvM,13516
-diffsynth_engine/models/qwen_image/qwen_image_vae.py,sha256=eO7f4YqiYXfw7NncBNFTu-xEvdJ5uKY-SnfP15QY0tE,38443
+diffsynth_engine/models/qwen_image/qwen_image_vae.py,sha256=FpauZV9IVvpvBeS9volu7kzH2mmCISS86AbHt0Jk2bQ,38442
 diffsynth_engine/models/sd/__init__.py,sha256=hjoKRnwoXOLD0wude-w7I6wK5ak7ACMbnbkPuBB2oU0,380
 diffsynth_engine/models/sd/sd_controlnet.py,sha256=kMGfIdriXhC7reT6iO2Z0rPICXEkXpytjeBQcR_sjT8,50577
 diffsynth_engine/models/sd/sd_text_encoder.py,sha256=BUOsBtSb7WH4Z37JhtYxOtpXMDJcQXZWzx_7JNbsJwM,5369
@@ -146,7 +146,7 @@ diffsynth_engine/pipelines/__init__.py,sha256=jh-4LSJ0vqlXiT8BgFgRIQxuAr2atEPyHr
 diffsynth_engine/pipelines/base.py,sha256=ShRiX5MY6bUkRKfuGrA1aalAqeHyeZxhzT87Mwc30b4,17231
 diffsynth_engine/pipelines/flux_image.py,sha256=L0ggxpthLD8a5-zdPHu9z668uWBei9YzPb4PFVypDNU,50707
 diffsynth_engine/pipelines/hunyuan3d_shape.py,sha256=TNV0Wr09Dj2bzzlpua9WioCClOj3YiLfE6utI9aWL8A,8164
-diffsynth_engine/pipelines/qwen_image.py,sha256=lrqwF3fikgQouifb-8KwWCxQhNVZard_7buoJqxHD7s,35759
+diffsynth_engine/pipelines/qwen_image.py,sha256=9n0fZCYw5E1iloXqd7vOU0XfHVPxQp_pm-v4D3Oloos,35751
 diffsynth_engine/pipelines/sd_image.py,sha256=nr-Nhsnomq8CsUqhTM3i2l2zG01YjwXdfRXgr_bC3F0,17891
 diffsynth_engine/pipelines/sdxl_image.py,sha256=v7ZACGPb6EcBunL6e5E9jynSQjE7GQx8etEV-ZLP91g,21704
 diffsynth_engine/pipelines/utils.py,sha256=HZbJHErNJS1DhlwJKvZ9dY7Kh8Zdlsw3zE2e88TYGRY,2277
@@ -190,8 +190,8 @@ diffsynth_engine/utils/video.py,sha256=8FCaeqIdUsWMgWI_6SO9SPynsToGcLCQAVYFTc4CD
 diffsynth_engine/utils/memory/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 diffsynth_engine/utils/memory/linear_regression.py,sha256=oW_EQEw13oPoyUrxiL8A7Ksa5AuJ2ynI2qhCbfAuZbg,3930
 diffsynth_engine/utils/memory/memory_predcit_model.py,sha256=EXprSl_zlVjgfMWNXP-iw83Ot3hyMcgYaRPv-dvyL84,3943
-diffsynth_engine-0.6.1.dev33.dist-info/licenses/LICENSE,sha256=x7aBqQuVI0IYnftgoTPI_A0I_rjdjPPQkjnU6N2nikM,11346
-diffsynth_engine-0.6.1.dev33.dist-info/METADATA,sha256=pgyNkuwU3lMQA66waiIU3BVtw-7zN3s8pEvinWC_LpI,1164
-diffsynth_engine-0.6.1.dev33.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-diffsynth_engine-0.6.1.dev33.dist-info/top_level.txt,sha256=6zgbiIzEHLbhgDKRyX0uBJOV3F6VnGGBRIQvSiYYn6w,17
-diffsynth_engine-0.6.1.dev33.dist-info/RECORD,,
+diffsynth_engine-0.6.1.dev34.dist-info/licenses/LICENSE,sha256=x7aBqQuVI0IYnftgoTPI_A0I_rjdjPPQkjnU6N2nikM,11346
+diffsynth_engine-0.6.1.dev34.dist-info/METADATA,sha256=Uu-yhnydrVudp5RdK0ifk9-q4J_18zulQge4fNs24Z0,1164
+diffsynth_engine-0.6.1.dev34.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+diffsynth_engine-0.6.1.dev34.dist-info/top_level.txt,sha256=6zgbiIzEHLbhgDKRyX0uBJOV3F6VnGGBRIQvSiYYn6w,17
+diffsynth_engine-0.6.1.dev34.dist-info/RECORD,,