Stable Diffusion model init for mlperf (#12314)

* include clip pr diff

* updated unet and sd init

* dehardcode default device

* revert beam hang workaround

---------

Co-authored-by: chenyu <chenyu@fastmail.com>

examples/mlperf/initializers.py

@@ -2,7 +2,9 @@ import math
 from typing import Union
 
 from tinygrad import Tensor, nn, dtypes
-from tinygrad.helpers import prod, argfix
+from tinygrad.helpers import prod, argfix, Context
+from tinygrad.nn.state import get_parameters
+from extra.models.unet import UNetModel
 
 # rejection sampling truncated randn
 def rand_truncn(*shape, dtype=None, truncstds=2, **kwargs) -> Tensor:
@@ -131,3 +133,59 @@ class Conv2dRetinaNet(nn.Conv2d):
   def __call__(self, x:Tensor) -> Tensor:
     return x.conv2d(self.weight.cast(dtypes.default_float), self.bias.cast(dtypes.default_float) if self.bias is not None else None,
                     groups=self.groups, stride=self.stride, dilation=self.dilation, padding=self.padding)
+
+# copy torch AMP: isolate mixed precision to just the below autocast ops, instead of using dtypes.default_float which affects all new Tensors
+class AutocastLinear(nn.Linear):
+  cast_dtype=dtypes.bfloat16 # enable monkeypatching of the mixed precision dtype
+  def __call__(self, x:Tensor) -> Tensor:
+    dtype = type(self).cast_dtype
+    return x.cast(dtype).linear(self.weight.cast(dtype).transpose(), self.bias.cast(dtype) if self.bias is not None else None)
+
+class AutocastConv2d(nn.Conv2d):
+  cast_dtype=dtypes.bfloat16
+  def __call__(self, x:Tensor) -> Tensor:
+    dtype = type(self).cast_dtype
+    return x.cast(dtype).conv2d(self.weight.cast(dtype), self.bias.cast(dtype), self.groups, self.stride, self.dilation, self.padding)
+
+# copy torch AMP: upcast to float32 before GroupNorm and LayerNorm
+class AutocastGroupNorm(nn.GroupNorm):
+  def __call__(self, x:Tensor) -> Tensor:
+    return super().__call__(x.cast(dtypes.float32))
+
+class AutocastLayerNorm(nn.LayerNorm):
+  def __call__(self, x:Tensor) -> Tensor:
+    return super().__call__(x.cast(dtypes.float32))
+
+def zero_module(module):
+  for p in get_parameters(module): p.assign(Tensor.zeros_like(p).contiguous())
+
+# Stable Diffusion mlperf reference doesn't call scaled_dot_product_attention
+# copy torch AMP: upcast to float32 before softmax on CUDA
+def attn_f32_softmax(q:Tensor, k:Tensor, v:Tensor) -> Tensor:
+  return (q.matmul(k.transpose(-2,-1), dtype=dtypes.float32) / math.sqrt(q.shape[-1])).softmax(-1).cast(q.dtype) @ v
+
+def init_stable_diffusion(version:str, pretrained:str, devices:list[str]):
+  from examples.stable_diffusion import StableDiffusion
+  from tinygrad.nn.state import safe_load, safe_save, load_state_dict, get_state_dict
+  from tempfile import TemporaryDirectory
+  model = StableDiffusion(version=version, pretrained=pretrained)
+  unet:UNetModel = model.model.diffusion_model
+
+  # this prevents extra consumption of memory, enabling much larger BS
+  Tensor.realize(*get_parameters(unet))
+  with TemporaryDirectory(prefix="unet_init") as tmp:
+    safe_save(get_state_dict(unet), init_fn:=f"{tmp}/init_model.safetensors")
+    load_state_dict(unet, safe_load(init_fn))
+
+  sqrt_alphas_cumprod = model.alphas_cumprod.sqrt().realize()
+  sqrt_one_minus_alphas_cumprod = (1 - model.alphas_cumprod).sqrt().realize()
+
+  if len(devices) > 1:
+    to_move = [sqrt_alphas_cumprod, sqrt_one_minus_alphas_cumprod]
+    if version == "v2-mlperf-train": to_move += get_parameters(unet) + get_parameters(model.cond_stage_model)
+    for p in to_move:
+      p.to_(devices)
+    with Context(BEAM=0):
+      Tensor.realize(*to_move)
+
+  return model, unet, sqrt_alphas_cumprod, sqrt_one_minus_alphas_cumprod

examples/stable_diffusion.py

@@ -9,11 +9,13 @@ from typing import Dict, Any
 from PIL import Image
 import numpy as np
 from tinygrad import Device, GlobalCounters, dtypes, Tensor, TinyJit
-from tinygrad.helpers import Timing, Context, getenv, fetch, colored, tqdm
+from tinygrad.helpers import Timing, Context, getenv, fetch, colored, tqdm, flatten
 from tinygrad.nn import Conv2d, GroupNorm
 from tinygrad.nn.state import torch_load, load_state_dict, get_state_dict
-from extra.models.clip import Closed, Tokenizer
+from extra.models.clip import Closed, Tokenizer, FrozenOpenClipEmbedder
+from extra.models import unet, clip
 from extra.models.unet import UNetModel
+from examples.mlperf.initializers import AutocastLinear, AutocastConv2d, AutocastGroupNorm, AutocastLayerNorm, zero_module, attn_f32_softmax, gelu_erf
 from extra.bench_log import BenchEvent, WallTimeEvent
 
 class AttnBlock:
@@ -154,12 +156,46 @@ unet_params: Dict[str,Any] = {
   "use_linear": False,
 }
 
+mlperf_params: Dict[str,Any] = {"adm_in_ch": None, "in_ch": 4, "out_ch": 4, "model_ch": 320, "attention_resolutions": [4, 2, 1], "num_res_blocks": 2,
+                                "channel_mult": [1, 2, 4, 4], "d_head": 64, "transformer_depth": [1, 1, 1, 1], "ctx_dim": 1024, "use_linear": True,
+                                "num_groups":16, "st_norm_eps":1e-6}
+
 class StableDiffusion:
-  def __init__(self):
+  def __init__(self, version:str|None=None, pretrained:str|None=None):
     self.alphas_cumprod = get_alphas_cumprod()
-    self.model = namedtuple("DiffusionModel", ["diffusion_model"])(diffusion_model = UNetModel(**unet_params))
-    self.first_stage_model = AutoencoderKL()
-    self.cond_stage_model = namedtuple("CondStageModel", ["transformer"])(transformer = namedtuple("Transformer", ["text_model"])(text_model = Closed.ClipTextTransformer()))
+    if version != "v2-mlperf-train":
+      self.first_stage_model = AutoencoderKL() # only needed for decoding generated latents to images; not needed in mlperf training from preprocessed moments
+
+    if not version:
+      self.cond_stage_model = namedtuple("CondStageModel", ["transformer"])(transformer = namedtuple("Transformer", ["text_model"])(text_model = Closed.ClipTextTransformer()))
+      unet_init_params = unet_params
+    elif version in {"v2-mlperf-train", "v2-mlperf-eval"}:
+      unet_init_params = mlperf_params
+      clip.gelu = gelu_erf
+      self.cond_stage_model = FrozenOpenClipEmbedder(**{"dims": 1024, "n_heads": 16, "layers": 24, "return_pooled": False, "ln_penultimate": True,
+                                                        "clip_tokenizer_version": "sd_mlperf_v5_0"})
+      unet.Linear, unet.Conv2d, unet.GroupNorm, unet.LayerNorm = AutocastLinear, AutocastConv2d, AutocastGroupNorm, AutocastLayerNorm
+      unet.attention, unet.gelu, unet.mixed_precision_dtype = attn_f32_softmax, gelu_erf, dtypes.bfloat16
+      if pretrained:
+        print("loading text encoder")
+        weights: dict[str,Tensor] = {k.replace("cond_stage_model.", "", 1):v for k,v in torch_load(pretrained)["state_dict"].items() if k.startswith("cond_stage_model.")}
+        weights["model.attn_mask"] = Tensor.full((77, 77), fill_value=float("-inf")).triu(1)
+        load_state_dict(self.cond_stage_model, weights)
+        # only the eval model needs the decoder
+        if version == "v2-mlperf-eval":
+          print("loading image latent encoder")
+          weights = {k.replace("first_stage_model.", "", 1):v for k,v in torch_load(pretrained)["state_dict"].items() if k.startswith("first_stage_model.")}
+          load_state_dict(self.first_stage_model, weights)
+
+    self.model = namedtuple("DiffusionModel", ["diffusion_model"])(diffusion_model = UNetModel(**unet_init_params))
+    if version == "v2-mlperf-train":
+      # the mlperf reference inits certain weights as zeroes
+      for bb in flatten(self.model.diffusion_model.input_blocks) + self.model.diffusion_model.middle_block + flatten(self.model.diffusion_model.output_blocks):
+        if isinstance(bb, unet.ResBlock):
+          zero_module(bb.out_layers[3])
+        elif isinstance(bb, unet.SpatialTransformer):
+          zero_module(bb.proj_out)
+      zero_module(self.model.diffusion_model.out[2])
 
   def get_x_prev_and_pred_x0(self, x, e_t, a_t, a_prev):
     temperature = 1