Update handler.py

handler.py

@@ -4,20 +4,45 @@ import os
 from typing import Any, Dict
 from PIL import Image
 import torch
+from diffusers import FluxPipeline
 import torch.distributed as dist
 from huggingface_inference_toolkit.logging import logger
-
+from para_attn.first_block_cache.diffusers_adapters import apply_cache_on_pipe
+# from torchao.quantization import quantize_, float8_dynamic_activation_float8_weight, float8_weight_only
+import time
 dist.init_process_group()
-torch.cuda.set_device(dist.get_rank())
 
-from para_attn.context_parallel import init_context_parallel_mesh
-from para_attn.context_parallel.diffusers_adapters import parallelize_pipe
-from para_attn.parallel_vae.diffusers_adapters import parallelize_vae
+torch.cuda.set_device(dist.get_rank())
 
 class EndpointHandler:
     def __init__(self,path=""):
 
-
+        self.pipe = FluxPipeline.from_pretrained(
+            "NoMoreCopyrightOrg/flux-dev",
+            torch_dtype=torch.bfloat16,
+        ).to("cuda")
+        from para_attn.context_parallel import init_context_parallel_mesh
+        from para_attn.context_parallel.diffusers_adapters import parallelize_pipe
+        from para_attn.parallel_vae.diffusers_adapters import parallelize_vae
+        mesh = init_context_parallel_mesh(
+            self.pipe.device.type,
+            max_ring_dim_size=2,
+        )
+        parallelize_pipe(
+            self.pipe,
+            mesh=mesh,
+        )
+        parallelize_vae(self.pipe.vae, mesh=mesh._flatten())
+        apply_cache_on_pipe(self.pipe, residual_diff_threshold=0.12)
+        # quantize_(self.pipe.text_encoder, float8_weight_only())
+        # quantize_(self.pipe.transformer, float8_dynamic_activation_float8_weight())
+        torch._inductor.config.reorder_for_compute_comm_overlap = True
+        self.pipe.transformer = torch.compile(
+            self.pipe.transformer, mode="max-autotune-no-cudagraphs",
+        )
+        self.pipe.vae = torch.compile(
+            self.pipe.vae, mode="max-autotune-no-cudagraphs",
+        )
 
     def __call__(self, data: Dict[str, Any]) -> Image.Image:
         logger.info(f"Received incoming request with {data=}")
@@ -32,4 +57,27 @@ class EndpointHandler:
                 " prompt to use for the image generation, and it needs to be a non-empty string."
             )
 
-        return "1"
+        parameters = data.pop("parameters", {})
+
+        num_inference_steps = parameters.get("num_inference_steps", 28)
+        width = parameters.get("width", 1024)
+        height = parameters.get("height", 1024)
+        guidance_scale = parameters.get("guidance_scale", 3.5)
+
+        # seed generator (seed cannot be provided as is but via a generator)
+        seed = parameters.get("seed", 0)
+        generator = torch.manual_seed(seed)
+        start_time = time.time()
+        result = self.pipe(  # type: ignore
+            prompt,
+            height=height,
+            width=width,
+            guidance_scale=guidance_scale,
+            num_inference_steps=num_inference_steps,
+            generator=generator,
+            # output_type="pil" if dist.get_rank() == 0 else "pt",
+        ).images[0]
+        end_time = time.time()
+        time_taken = end_time - start_time
+        print(f"Time taken: {time_taken:.2f} seconds")
+        return result