precompile for diffusion (#860)

Co-authored-by: whz <wuhuanzi@wuhuanziMacBook-Air.local>

Author: pathfinder-pf

python/pyproject.toml

@@ -12,7 +12,7 @@ license = { file = "LICENSE" }
 
 dependencies = [
     "fastapi~=0.116.1",
-    "flax~=0.12.0",
+    "flax==0.12.4",
     "huggingface-hub~=0.34.3",
     "jinja2~=3.1.6",
     "llguidance~=1.3.0",

python/sgl_jax/srt/multimodal/common/ServerArgs.py

@@ -19,8 +19,8 @@ class MultimodalServerArgs(ServerArgs):
     text_encoder_precisions: tuple[str, ...] = field(default_factory=lambda: ("fp32",))
     image_encoder_precision: str = "bf16"
 
-    vae_decode_precompile_width_height: list[str] | None = None
-    vae_decode_precompile_frame_paddings: list[int] | None = None
+    precompile_width_heights: list[str] | None = None
+    precompile_frame_paddings: list[int] | None = None
 
     @staticmethod
     def add_cli_args(parser: argparse.ArgumentParser):
@@ -95,14 +95,14 @@ def add_cli_args(parser: argparse.ArgumentParser):
         )
 
         parser.add_argument(
-            "--vae-decode-precompile-width-height",
+            "--precompile-width-heights",
             type=str,
             nargs="+",
             help="Set the list of width and height for jax jit, format width*height",
         )
 
         parser.add_argument(
-            "--vae-decode-precompile-frame-paddings",
+            "--precompile-frame-paddings",
             type=int,
             nargs="+",
             help="Set the frame count list for jax jit",
@@ -115,17 +115,17 @@ def __post_init__(self):
         # manually.
         super().__post_init__()
 
-        if self.vae_decode_precompile_width_height is not None:
-            for wh in self.vae_decode_precompile_width_height:
+        if self.precompile_width_heights is not None:
+            for wh in self.precompile_width_heights:
                 if len(wh.split("*")) < 2:
                     raise Exception("Width and height must be connected with an asterisk *.")
-            if self.vae_decode_precompile_frame_paddings is None:
-                self.vae_decode_precompile_frame_paddings = [1]
+            if self.precompile_frame_paddings is None:
+                self.precompile_frame_paddings = [1]
             else:
-                self.vae_decode_precompile_frame_paddings.sort()
+                self.precompile_frame_paddings.sort()
         else:
-            self.vae_decode_precompile_width_height = ["480*832"]
-            self.vae_decode_precompile_frame_paddings = [1]
+            self.precompile_width_heights = ["480*832"]
+            self.precompile_frame_paddings = [1]
 
     @classmethod
     def from_cli_args(cls, args: argparse.Namespace):

python/sgl_jax/srt/multimodal/entrypoint/http_server.py

@@ -554,14 +554,6 @@ def _execute_multimodal_server_warmup(
             ],
             "max_tokens": 3,
         }
-    elif _is_wan_model(server_args.model_path):
-        request_endpoint = "/api/v1/images/generation"
-        json_data = {
-            "prompt": "warmup request",
-            "size": "480*832",
-            "num_inference_steps": 2,
-            "save_output": False,
-        }
     elif "MiMo-Audio" in server_args.model_path:
         request_endpoint = "/v1/audio/transcriptions"
         # audio_url = "https://huggingface.co/datasets/nvidia/AudioSkills/resolve/main/assets/WhDJDIviAOg_120_10.mp3"

python/sgl_jax/srt/multimodal/manager/scheduler/diffusion_scheduler.py

@@ -59,6 +59,10 @@ def __init__(
         self.aborted_rids: set[str] = set()
         # Current request being processed (for abort checking during steps)
         self._current_rid: str | None = None
+        if not server_args.disable_precompile:
+            logger.info("[Diffusion Scheduler] Begins to run diffusion worker precompile.")
+            self.diffusion_worker.run_precompile()
+            logger.info("[Diffusion Scheduler] Completes diffusion worker precompile.")
 
     def event_loop_normal(self):
         """Blocking event loop for processing incoming diffusion requests.

python/sgl_jax/srt/multimodal/manager/scheduler/vae_scheduler.py

@@ -127,10 +127,10 @@ def preprocess(self, req):
             req.latents += self.model_config.shift_factor
         req.latents = jax.device_get(req.latents)
         latents_t_padding = 0
-        if self.server_args.vae_decode_precompile_frame_paddings is not None and hasattr(
+        if self.server_args.precompile_frame_paddings is not None and hasattr(
             self.model_config, "scale_factor_temporal"
         ):
-            for n_frame in self.server_args.vae_decode_precompile_frame_paddings:
+            for n_frame in self.server_args.precompile_frame_paddings:
                 latents_t = (n_frame - 1) // self.model_config.scale_factor_temporal + 1
                 if latents_t >= req.latents.shape[1]:
                     latents_t_padding = latents_t - req.latents.shape[1]
@@ -153,7 +153,8 @@ def run_vae_batch(self, batch: list[Req]):
 
         for req in batch:
             output, cache_miss = self.vae_worker.forward(req)
-            logger.info("VAE forward pass cache miss: %s", cache_miss)
+            if cache_miss > 0:
+                logger.info("VAE forward pass cache miss: %s", cache_miss)
             req.output = jax.device_get(output[:, : req.num_frames, :, :, :])
             req.latents = None
             self.forward_ct += 1

python/sgl_jax/srt/multimodal/model_executor/diffusion/diffusion_model_runner.py

@@ -222,7 +222,8 @@ def pad_to_512(embeds):
                     encoder_hidden_states_image=None,
                     guidance_scale=None,
                 )
-                logger.info("diffusion cache miss count: %d", count())
+                if count() > 0:
+                    logger.info("diffusion cache miss count: %d", count())
             if do_classifier_free_guidance:
                 bsz = latents.shape[0] // 2
                 noise_uncond = noise_pred[bsz:]

python/sgl_jax/srt/multimodal/model_executor/diffusion/diffusion_model_worker.py

@@ -1,12 +1,22 @@
+import logging
+import time
 from collections.abc import Callable
 
 import jax
+import numpy as np
+from jax.sharding import NamedSharding
+from jax.sharding import PartitionSpec as P
+from tqdm import tqdm
 
 from sgl_jax.srt.multimodal.common.ServerArgs import MultimodalServerArgs
+from sgl_jax.srt.multimodal.configs.config_registry import get_diffusion_config
 from sgl_jax.srt.multimodal.manager.schedule_batch import Req
 from sgl_jax.srt.multimodal.model_executor.diffusion.diffusion_model_runner import (
     DiffusionModelRunner,
 )
+from sgl_jax.srt.utils.jax_utils import device_array
+
+logger = logging.getLogger(__name__)
 
 
 class DiffusionModelWorker:
@@ -22,6 +32,9 @@ def __init__(
             server_args, self.mesh, model_class=model_class, stage_sub_dir=stage_sub_dir
         )
         self.initialize()
+        self.precompile_width_heights = server_args.precompile_width_heights
+        self.precompile_frame_paddings = server_args.precompile_frame_paddings
+        self.model_config = get_diffusion_config(server_args.model_path)
 
     def initialize(self):
         pass
@@ -50,3 +63,38 @@ def forward(
         return self.model_runner.forward(
             batch, mesh, abort_checker=abort_checker, step_callback=step_callback
         )
+
+    def run_precompile(self):
+        self.precompile()
+
+    def precompile(self):
+        start_time = time.perf_counter()
+        logger.info(
+            "[DIFFUSION] Begin to precompile width*height=%s",
+            self.precompile_width_heights,
+        )
+
+        with tqdm(
+            self.precompile_width_heights, desc="[DIFFUSION] PRECOMPILE", leave=False
+        ) as pbar:
+            for wh in pbar:
+                whs = wh.split("*")
+                width, height = int(whs[0]), int(whs[1])
+                assert width % self.model_config.scale_factor_spatial == 0
+                assert height % self.model_config.scale_factor_spatial == 0
+                for t in self.precompile_frame_paddings:
+                    pbar.set_postfix(wh=wh, t=t)
+                    embeds = np.random.random((2, 512, self.model_config.text_dim))
+                    embeds = device_array(embeds, sharding=NamedSharding(self.mesh, P()))
+                    req = Req(
+                        prompt_embeds=embeds[0],
+                        negative_prompt_embeds=embeds[1],
+                        do_classifier_free_guidance=True,
+                        width=width,
+                        height=height,
+                        num_frames=t,
+                        num_inference_steps=1,
+                    )
+                    self.model_runner.forward(req, self.mesh)
+        end_time = time.perf_counter()
+        logger.info("[DIFFUSION] Precompile finished in %.0f secs", end_time - start_time)

python/sgl_jax/srt/multimodal/model_executor/vae/vae_model_worker.py

@@ -28,8 +28,8 @@ def __init__(
             server_args, mesh, model_class=model_class, stage_sub_dir=stage_sub_dir
         )
         self.server_args = server_args
-        self.vae_decode_precompile_width_height = server_args.vae_decode_precompile_width_height
-        self.vae_decode_precompile_frame_paddings = server_args.vae_decode_precompile_frame_paddings
+        self.precompile_width_heights = server_args.precompile_width_heights
+        self.precompile_frame_paddings = server_args.precompile_frame_paddings
         self.model_config = get_vae_config(self.server_args.model_path)
         # Initialize model here based on model_config
 
@@ -46,23 +46,23 @@ def decode_precompile(self):
         start_time = time.perf_counter()
         logger.info(
             "[VAE DECODE] Begin to precompile width*height=%s",
-            self.vae_decode_precompile_width_height,
+            self.precompile_width_heights,
         )
 
         with tqdm(
-            self.vae_decode_precompile_width_height, desc="[VAE DECODE] PRECOMPILE", leave=False
+            self.precompile_width_heights, desc="[VAE DECODE] PRECOMPILE", leave=False
         ) as pbar:
             for wh in pbar:
                 whs = wh.split("*")
                 width, height = int(whs[0]), int(whs[1])
                 assert width % self.model_config.scale_factor_spatial == 0
                 assert height % self.model_config.scale_factor_spatial == 0
-                for t in self.vae_decode_precompile_frame_paddings:
+                for t in self.precompile_frame_paddings:
                     pbar.set_postfix(wh=wh, t=t)
                     latents_cpu = np.random.random(
                         (
                             1,
-                            t // self.model_config.scale_factor_temporal + 1,
+                            (t - 1) // self.model_config.scale_factor_temporal + 1,
                             height // self.model_config.scale_factor_spatial,
                             width // self.model_config.scale_factor_spatial,
                             self.model_config.z_dim,