Upload 2 files

handler.py

@@ -20,14 +20,14 @@ from huggingface_inference_toolkit.logging import logger
 import subprocess
 subprocess.run("pip list", shell=True)
 
-print("device name:", torch.cuda.get_device_name())
-print("device capability:", torch.cuda.get_device_capability())
+print("Device Name:", torch.cuda.get_device_name())
+print("Device Capability:", torch.cuda.get_device_capability())
 
 IS_TURBO = False
-IS_4BIT = True
+IS_4BIT = False
 IS_PARA = True
 IS_LVRAM = False
-IS_COMPILE = False
+IS_COMPILE = True
 IS_AUTOQ = False
 IS_CC90 = True if torch.cuda.get_device_capability() >= (9, 0) else False
 IS_CC89 = True if torch.cuda.get_device_capability() >= (8, 9) else False
@@ -160,7 +160,7 @@ def load_pipeline_opt(repo_id: str, dtype: torch.dtype) -> Any:
     else: quantize_(vae, weight, device="cuda")
     vae.to(memory_format=torch.channels_last)
     vae = torch.compile(vae, mode="max-autotune", fullgraph=True)
-    pipe = FluxPipeline.from_pretrained(repo_id, transformer=None, vae=None, text_encoder_2=load_te2(repo_id, dtype), torch_dtype=dtype, quantization_config=quantization_config)
+    pipe = FluxPipeline.from_pretrained(repo_id, transformer=None, vae=None, torch_dtype=dtype, quantization_config=quantization_config)
     pipe.transformer = transformer
     pipe.vae = vae
     return pipe

requirements.txt

@@ -15,6 +15,7 @@ sentencepiece
 protobuf
 triton
 gemlite
+tabulate
 para-attn
 bitsandbytes
 optimum-quanto