Spaces:
Running
on
Zero
Running
on
Zero
Update inferencer.py
Browse files- inferencer.py +48 -50
inferencer.py
CHANGED
@@ -233,57 +233,55 @@ class InterleaveInferencer:
|
|
233 |
|
234 |
current_image_shapes = image_shapes
|
235 |
|
236 |
-
|
237 |
-
|
238 |
-
# with torch.autocast(device_type="cuda", enabled=True, dtype=torch.bfloat16):
|
239 |
-
|
240 |
-
if think:
|
241 |
-
system_prompt = VLM_THINK_SYSTEM_PROMPT if understanding_output else GEN_THINK_SYSTEM_PROMPT
|
242 |
-
gen_context = self.update_context_text(system_prompt, gen_context)
|
243 |
-
cfg_text_context = self.update_context_text(system_prompt, cfg_text_context)
|
244 |
-
cfg_img_context = self.update_context_text(system_prompt, cfg_img_context)
|
245 |
-
|
246 |
-
for input_term in input_lists:
|
247 |
-
if isinstance(input_term, str):
|
248 |
-
gen_context = self.update_context_text(input_term, gen_context)
|
249 |
-
cfg_text_context = self.update_context_text(input_term, cfg_text_context)
|
250 |
-
cfg_img_context = self.update_context_text(input_term, cfg_img_context)
|
251 |
-
elif isinstance(input_term, Image.Image):
|
252 |
-
current_image_shapes = input_term.size[::-1] # H, W
|
253 |
-
use_vae_for_input_image = not understanding_output
|
254 |
-
gen_context = self.update_context_image(input_term, gen_context, vae=use_vae_for_input_image, vit=True)
|
255 |
-
cfg_text_context = self.update_context_image(input_term, cfg_text_context, vae=use_vae_for_input_image, vit=True)
|
256 |
-
# cfg_img_context does not typically see input images
|
257 |
-
else:
|
258 |
-
raise ValueError(f"Unsupported input type: {type(input_term)}")
|
259 |
-
|
260 |
-
if understanding_output: # Generate text
|
261 |
-
yield from self.gen_text(gen_context, max_length=max_think_token_n, do_sample=do_sample, temperature=temperature)
|
262 |
-
else: # Generate image
|
263 |
if think:
|
264 |
-
|
265 |
-
|
266 |
-
|
267 |
-
|
268 |
-
|
269 |
-
|
270 |
-
|
271 |
-
|
272 |
-
|
273 |
-
|
274 |
-
|
275 |
-
|
276 |
-
|
277 |
-
|
278 |
-
|
279 |
-
|
280 |
-
|
281 |
-
|
282 |
-
|
283 |
-
|
284 |
-
|
285 |
-
|
286 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
287 |
|
288 |
def __call__(
|
289 |
self,
|
|
|
233 |
|
234 |
current_image_shapes = image_shapes
|
235 |
|
236 |
+
with torch.autocast(device_type="cuda", enabled=True, dtype=torch.bfloat16):
|
237 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
238 |
if think:
|
239 |
+
system_prompt = VLM_THINK_SYSTEM_PROMPT if understanding_output else GEN_THINK_SYSTEM_PROMPT
|
240 |
+
gen_context = self.update_context_text(system_prompt, gen_context)
|
241 |
+
cfg_text_context = self.update_context_text(system_prompt, cfg_text_context)
|
242 |
+
cfg_img_context = self.update_context_text(system_prompt, cfg_img_context)
|
243 |
+
|
244 |
+
for input_term in input_lists:
|
245 |
+
if isinstance(input_term, str):
|
246 |
+
gen_context = self.update_context_text(input_term, gen_context)
|
247 |
+
cfg_text_context = self.update_context_text(input_term, cfg_text_context)
|
248 |
+
cfg_img_context = self.update_context_text(input_term, cfg_img_context)
|
249 |
+
elif isinstance(input_term, Image.Image):
|
250 |
+
current_image_shapes = input_term.size[::-1] # H, W
|
251 |
+
use_vae_for_input_image = not understanding_output
|
252 |
+
gen_context = self.update_context_image(input_term, gen_context, vae=use_vae_for_input_image, vit=True)
|
253 |
+
cfg_text_context = self.update_context_image(input_term, cfg_text_context, vae=use_vae_for_input_image, vit=True)
|
254 |
+
# cfg_img_context does not typically see input images
|
255 |
+
else:
|
256 |
+
raise ValueError(f"Unsupported input type: {type(input_term)}")
|
257 |
+
|
258 |
+
if understanding_output: # Generate text
|
259 |
+
yield from self.gen_text(gen_context, max_length=max_think_token_n, do_sample=do_sample, temperature=temperature)
|
260 |
+
else: # Generate image
|
261 |
+
if think:
|
262 |
+
thought_text_parts = []
|
263 |
+
for part in self.gen_text(gen_context, max_length=max_think_token_n, do_sample=do_sample, temperature=temperature):
|
264 |
+
yield part # Stream the thought
|
265 |
+
thought_text_parts.append(part)
|
266 |
+
full_thought_text = "".join(thought_text_parts)
|
267 |
+
if full_thought_text: # Only update if thought was generated
|
268 |
+
gen_context = self.update_context_text(full_thought_text, gen_context)
|
269 |
+
cfg_text_context = self.update_context_text(full_thought_text, cfg_text_context)
|
270 |
+
|
271 |
+
img = self.gen_image(
|
272 |
+
image_shape=current_image_shapes,
|
273 |
+
gen_context=gen_context,
|
274 |
+
cfg_text_precontext=cfg_text_context,
|
275 |
+
cfg_img_precontext=cfg_img_context,
|
276 |
+
cfg_text_scale=cfg_text_scale,
|
277 |
+
cfg_img_scale=cfg_img_scale,
|
278 |
+
cfg_interval=cfg_interval,
|
279 |
+
timestep_shift=timestep_shift,
|
280 |
+
num_timesteps=num_timesteps,
|
281 |
+
cfg_renorm_min=cfg_renorm_min,
|
282 |
+
cfg_renorm_type=cfg_renorm_type,
|
283 |
+
)
|
284 |
+
yield img
|
285 |
|
286 |
def __call__(
|
287 |
self,
|