[RFC] Clean and refactor TeaCache registry (#416)
Signed-off-by: yuanheng <jonathan.zhaoyh@gmail.com>
This commit is contained in:
1 parent
a16476bebf
commit
edbd610060
9 files changed
+89
-87
No files matched your search
@@ -230,9 +230,6 @@ class QwenImageEditPipeline(
|
||||
model, subfolder="processor", local_files_only=local_files_only
|
||||
)
|
||||
|
||||
# Initialize cache backend to None (will be set by worker if needed)
|
||||
self._cache_backend = None
|
||||
|
||||
self.stage = None
|
||||
|
||||
self.vae_scale_factor = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8
|
||||
|
||||
@@ -194,7 +194,6 @@ class QwenImageEditPlusPipeline(
|
||||
)
|
||||
|
||||
self.stage = None
|
||||
self._cache_backend = None
|
||||
|
||||
self.vae_scale_factor = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8
|
||||
self.latent_channels = self.vae.config.z_dim if getattr(self, "vae", None) else 16
|
||||
|
||||
@@ -264,8 +264,6 @@ the image\n<|vision_start|><|image_pad|><|vision_end|><|im_end|>\n<|im_start|>as
|
||||
)
|
||||
self.default_sample_size = 128
|
||||
|
||||
# VLLM-omni related, Initialize cache backend to None (will be set by worker if needed)
|
||||
self._cache_backend = None
|
||||
self.stage = None
|
||||
|
||||
def check_inputs(
|
||||
@@ -570,42 +568,36 @@ the image\n<|vision_start|><|image_pad|><|vision_end|><|im_end|>\n<|im_start|>as
|
||||
|
||||
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
|
||||
timestep = t.expand(latents.shape[0]).to(latents.dtype)
|
||||
# Forward pass for positive prompt (or unconditional if no CFG)
|
||||
# cache_branch is passed to hook for CFG-aware state management
|
||||
transformer_kwargs = {
|
||||
"hidden_states": latent_model_input,
|
||||
"timestep": timestep / 1000,
|
||||
"guidance": guidance,
|
||||
"encoder_hidden_states_mask": prompt_embeds_mask,
|
||||
"encoder_hidden_states": prompt_embeds,
|
||||
"img_shapes": img_shapes,
|
||||
"txt_seq_lens": txt_seq_lens,
|
||||
"attention_kwargs": self.attention_kwargs,
|
||||
"additional_t_cond": is_rgb,
|
||||
"return_dict": False,
|
||||
}
|
||||
if self._cache_backend is not None:
|
||||
transformer_kwargs["cache_branch"] = "positive"
|
||||
|
||||
noise_pred = self.transformer(**transformer_kwargs)[0]
|
||||
self.transformer.do_true_cfg = do_true_cfg # used in teacache hook
|
||||
# Forward pass for positive prompt (or unconditional if no CFG)
|
||||
noise_pred = self.transformer(
|
||||
hidden_states=latent_model_input,
|
||||
timestep=timestep / 1000,
|
||||
guidance=guidance,
|
||||
encoder_hidden_states_mask=prompt_embeds_mask,
|
||||
encoder_hidden_states=prompt_embeds,
|
||||
img_shapes=img_shapes,
|
||||
txt_seq_lens=txt_seq_lens,
|
||||
attention_kwargs=self.attention_kwargs,
|
||||
additional_t_cond=is_rgb,
|
||||
return_dict=False,
|
||||
)[0]
|
||||
noise_pred = noise_pred[:, : latents.size(1)]
|
||||
|
||||
if do_true_cfg:
|
||||
neg_transformer_kwargs = {
|
||||
"hidden_states": latent_model_input,
|
||||
"timestep": timestep / 1000,
|
||||
"guidance": guidance,
|
||||
"encoder_hidden_states_mask": negative_prompt_embeds_mask,
|
||||
"encoder_hidden_states": negative_prompt_embeds,
|
||||
"img_shapes": img_shapes,
|
||||
"txt_seq_lens": negative_txt_seq_lens,
|
||||
"attention_kwargs": self.attention_kwargs,
|
||||
"additional_t_cond": is_rgb,
|
||||
"return_dict": False,
|
||||
}
|
||||
if self._cache_backend is not None:
|
||||
neg_transformer_kwargs["cache_branch"] = "negative"
|
||||
neg_noise_pred = self.transformer(**neg_transformer_kwargs)[0]
|
||||
neg_noise_pred = self.transformer(
|
||||
hidden_states=latent_model_input,
|
||||
timestep=timestep / 1000,
|
||||
guidance=guidance,
|
||||
encoder_hidden_states_mask=negative_prompt_embeds_mask,
|
||||
encoder_hidden_states=negative_prompt_embeds,
|
||||
img_shapes=img_shapes,
|
||||
txt_seq_lens=negative_txt_seq_lens,
|
||||
attention_kwargs=self.attention_kwargs,
|
||||
additional_t_cond=is_rgb,
|
||||
return_dict=False,
|
||||
)[0]
|
||||
|
||||
neg_noise_pred = neg_noise_pred[:, : latents.size(1)]
|
||||
comb_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred)
|
||||
|
||||
Reference in new issue
Block a user