chore: t5 and clip are flax models now

jflux/cli.py

@@ -3,7 +3,6 @@
 import time
 from dataclasses import dataclass
 from glob import iglob
-import torch
 import jax
 import jax.numpy as jnp
 import numpy as np
@@ -14,11 +13,13 @@
 from jflux.sampling import denoise, get_noise, get_schedule, prepare, unpack
 from jflux.util import configs, load_ae, load_clip, load_flow_model, load_t5
 
+
 def torch2jax(tensor):
     tensor = tensor.float().numpy()
     tensor = jnp.array(tensor, dtype=jnp.bfloat16)
     return tensor
 
+
 @dataclass
 class SamplingOptions:
     prompt: str

jflux/modules/conditioner.py

@@ -1,51 +1,49 @@
-# Note: This is a torch module not a Jax module
-import jax.numpy as jnp
+from flax import nnx
 from chex import Array
-from torch import nn
-from transformers import CLIPTextModel, CLIPTokenizer, T5EncoderModel, T5Tokenizer
+from transformers import (
+    FlaxCLIPTextModel,
+    CLIPTokenizer,
+    FlaxT5EncoderModel,
+    T5Tokenizer,
+)
 
 
-class HFEmbedder(nn.Module):
+class HFEmbedder(nnx.Module):
     def __init__(self, version: str, max_length: int, **hf_kwargs):
-        super().__init__()
-        self.is_clip = version.startswith("openai")
+        self.is_clip = version.split("/")[1].startswith("clip")
         self.max_length = max_length
         self.output_key = "pooler_output" if self.is_clip else "last_hidden_state"
 
         if self.is_clip:
             self.tokenizer: CLIPTokenizer = CLIPTokenizer.from_pretrained(
                 version, max_length=max_length
             )
-            self.hf_module: CLIPTextModel = CLIPTextModel.from_pretrained(
+            self.hf_module: FlaxCLIPTextModel = FlaxCLIPTextModel.from_pretrained(
                 version, **hf_kwargs
             )
         else:
             self.tokenizer: T5Tokenizer = T5Tokenizer.from_pretrained(
                 version, max_length=max_length
             )
-            self.hf_module: T5EncoderModel = T5EncoderModel.from_pretrained(
+            self.hf_module: FlaxT5EncoderModel = FlaxT5EncoderModel.from_pretrained(
                 version, **hf_kwargs
             )
 
-        self.hf_module = self.hf_module.eval().requires_grad_(False)
-
-    def forward(self, text: list[str]) -> Array:
+    def __call__(self, text: list[str]) -> Array:
         batch_encoding = self.tokenizer(
             text,
             truncation=True,
             max_length=self.max_length,
             return_length=False,
             return_overflowing_tokens=False,
             padding="max_length",
-            return_tensors="pt",
+            return_tensors="np",
         )
 
         outputs = self.hf_module(
-            input_ids=batch_encoding["input_ids"].to(self.hf_module.device),
+            input_ids=batch_encoding["input_ids"],
             attention_mask=None,
             output_hidden_states=False,
         )
-        torch_outputs = outputs[self.output_key]
-
-        jax_outputs = jnp.array(torch_outputs.cpu().float(), dtype=jnp.bfloat16)
-        return jax_outputs
+        outputs = outputs[self.output_key]
+        return outputs

jflux/util.py

@@ -1,7 +1,7 @@
 import os
 from dataclasses import dataclass
 
-import torch  # need for t5 and clip
+import torch  # need for torch 2 jax
 from flax import nnx
 from huggingface_hub import hf_hub_download
 import jax
@@ -13,6 +13,7 @@
 from jflux.modules.conditioner import HFEmbedder
 from jflux.port import port_autoencoder, port_flux
 
+
 def torch2jax(torch_tensor):
     intermediate_tensor = torch_tensor.to(torch.float32)
     jax_tensor = jnp.array(intermediate_tensor, dtype=jnp.bfloat16)
@@ -144,17 +145,17 @@ def load_flow_model(name: str, hf_download: bool = True) -> Flux:
 
 
 def load_t5() -> HFEmbedder:
-    device = "cuda" if torch.cuda.is_available() else "cpu"
     return HFEmbedder(
-        "google/t5-v1_1-xxl", max_length=512, torch_dtype=torch.bfloat16
-    ).to(device)
+        "ariG23498/t5-v1-1-xxl-flax",
+        max_length=512,
+    )
 
 
 def load_clip() -> HFEmbedder:
-    device = "cuda" if torch.cuda.is_available() else "cpu"
     return HFEmbedder(
-        "openai/clip-vit-large-patch14", max_length=77, torch_dtype=torch.bfloat16
-    ).to(device)
+        "ariG23498/clip-vit-large-patch14-text-flax",
+        max_length=77,
+    )
 
 
 def load_ae(name: str, hf_download: bool = True) -> AutoEncoder: