File size: 9,078 Bytes
4900749 046a4b1 4900749 046a4b1 4900749 046a4b1 4900749 046a4b1 4900749 046a4b1 4900749 046a4b1 4900749 046a4b1 4900749 046a4b1 4900749 046a4b1 4900749 046a4b1 4900749 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 | # Copyright 2025 Dhruv Nair. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Denoising loop for RFDiffusion3.
Implements the iterative denoising procedure from the original
inference_sampler.py (SampleDiffusionWithMotif.sample_diffusion_like_af3).
The loop iterates over consecutive pairs (c_t_minus_1, c_t) in the noise schedule:
1. Inject stochastic noise: t_hat = c_t_minus_1 * (gamma + 1), epsilon ~ N(0, noise_scale * sqrt(t_hat^2 - c_t_minus_1^2))
2. Call model: X_denoised = model(X_noisy, t_hat)
3. Euler update: X = X_noisy + step_scale * (c_t - t_hat) * (X_noisy - X_denoised) / t_hat
"""
from typing import Callable, List
import torch
from diffusers.utils import logging
from diffusers.modular_pipelines import ModularPipeline, ModularPipelineBlocks, PipelineState
from diffusers.modular_pipelines.modular_pipeline_utils import ComponentSpec, InputParam, OutputParam
logger = logging.get_logger(__name__)
class RFDiffusionDenoiseStep(ModularPipelineBlocks):
"""
Iterative denoising step for RFDiffusion3.
Implements the EDM stochastic sampling loop matching the original
SampleDiffusionWithMotif.sample_diffusion_like_af3.
"""
model_name = "rfdiffusion"
@property
def description(self) -> str:
return (
"Iteratively denoise protein structure through reverse diffusion. "
"Uses EDM stochastic sampling with gamma noise injection and step scaling."
)
@property
def expected_components(self) -> List[ComponentSpec]:
return [
ComponentSpec("transformer", description="RFDiffusion transformer for structure prediction"),
ComponentSpec("scheduler", description="Scheduler for noise injection and stepping"),
]
@property
def inputs(self) -> List[InputParam]:
return [
InputParam(
"n_recycle",
default=None,
type_hint=int,
description="Number of recycling iterations (None uses model default)",
),
InputParam(
"callback",
type_hint=Callable,
description="Optional callback function called at each step",
),
InputParam(
"callback_steps",
default=1,
type_hint=int,
description="Frequency of callback invocation",
),
InputParam("xyz", required=True, type_hint=torch.Tensor, description="Initial noised coordinates [D, L_atom, 3]"),
InputParam("noise_schedule", required=True, type_hint=torch.Tensor, description="EDM noise schedule"),
InputParam("motif_mask", required=True, type_hint=torch.Tensor, description="Mask for fixed motif positions"),
InputParam("f", required=True, type_hint=dict, description="Foundry feature dict"),
InputParam("initializer_outputs", required=True, type_hint=dict, description="Embedded conditioning"),
]
@property
def intermediate_outputs(self) -> List[OutputParam]:
return [
OutputParam("xyz", type_hint=torch.Tensor, description="Denoised coordinates [D, L, 3]"),
OutputParam("xyz_backbone", type_hint=torch.Tensor, description="Backbone N/CA/C/O coordinates [D, L, 4, 3]"),
OutputParam("single", type_hint=torch.Tensor, description="Single representation"),
OutputParam("pair", type_hint=torch.Tensor, description="Pair representation"),
OutputParam("sequence_logits", type_hint=torch.Tensor, description="Predicted sequence logits"),
OutputParam("sequence_indices", type_hint=torch.Tensor, description="Predicted sequence indices"),
OutputParam("trajectory", type_hint=List[torch.Tensor], description="Denoising trajectory"),
]
@torch.no_grad()
def __call__(self, components: ModularPipeline, state: PipelineState) -> PipelineState:
# Both components are required. Falling back to a hand-rolled Euler step when the
# scheduler is missing silently swaps out the EDM sampler and yields coordinates that
# look finite but are not a protein backbone.
if components.transformer is None:
raise ValueError(
"`transformer` is not loaded. Call `load_components(trust_remote_code=True)` on the "
"pipeline before calling it."
)
if components.scheduler is None:
raise ValueError(
"`scheduler` is not loaded. Call `load_components(trust_remote_code=True)` on the "
"pipeline before calling it."
)
block_state = self.get_block_state(state)
xyz = block_state.xyz
noise_schedule = block_state.noise_schedule
motif_mask = block_state.motif_mask
f = block_state.f
initializer_outputs = block_state.initializer_outputs
n_recycle = block_state.n_recycle
callback = block_state.callback
callback_steps = block_state.callback_steps or 1
X_denoised_L_traj = []
X_L = xyz.clone()
D = X_L.shape[0]
device = X_L.device
# Ensure all tensors are on the same device as xyz
noise_schedule = noise_schedule.to(device)
if motif_mask is not None:
motif_mask = motif_mask.to(device)
single = None
pair = None
sequence_logits = None
sequence_indices = None
# Iterate over consecutive pairs (c_t_minus_1, c_t) in the noise schedule
# noise_schedule goes from high noise to low noise
for step_num in range(len(noise_schedule) - 1):
c_t_minus_1 = noise_schedule[step_num]
c_t = noise_schedule[step_num + 1]
# Step 1: Inject stochastic noise (matching original sampler)
X_noisy_L, t_hat = components.scheduler.add_noise(
X_L, c_t_minus_1, c_t, motif_mask=motif_mask
)
# Step 2: Model forward pass
# t_hat is a scalar, tile to batch dimension
t_batch = (t_hat.to(device).expand(D) if isinstance(t_hat, torch.Tensor)
else torch.full((D,), t_hat, device=device))
output = components.transformer(
xyz_noisy=X_noisy_L,
t=t_batch,
f=f,
initializer_outputs=initializer_outputs,
n_recycle=n_recycle,
)
X_denoised_L = output.xyz
single = output.single
pair = output.pair
sequence_logits = output.sequence_logits
sequence_indices = output.sequence_indices
# Step 3: Euler update with step_scale (matching original sampler)
X_L = components.scheduler.step(
xyz_pred=X_denoised_L,
xyz_noisy=X_noisy_L,
c_t_minus_1=c_t_minus_1,
c_t=c_t,
motif_mask=motif_mask,
)
X_denoised_L_traj.append(X_denoised_L.clone())
if callback is not None and step_num % callback_steps == 0:
callback(step_num, c_t_minus_1, X_L)
# The sampler runs on padded atom-level coordinates (14 slots per residue). Downstream
# blocks and the documented output are one point per residue, so collapse to CA here.
is_ca = f["is_ca"]
block_state.xyz = X_L[:, is_ca]
# MPNN wants N, CA, C and O per residue. They are present in the padded tensor, so select
# them instead of synthesising them from CA: foundry's `BACKBONE_ATOM_NAMES` is
# ["N", "CA", "C", "O"], which is also the CCD atom order for amino acids.
is_backbone = f["is_backbone"]
n_residues = int(is_ca.sum())
n_backbone = int(is_backbone.sum())
if n_backbone != 4 * n_residues:
raise ValueError(
f"Expected 4 backbone atoms per residue, got {n_backbone} for {n_residues} residues. "
"The padded atom layout does not match the N/CA/C/O ordering MPNN assumes."
)
block_state.xyz_backbone = X_L[:, is_backbone].reshape(X_L.shape[0], n_residues, 4, 3)
block_state.single = single
block_state.pair = pair
block_state.sequence_logits = sequence_logits
block_state.sequence_indices = sequence_indices
block_state.trajectory = [step[:, is_ca] for step in X_denoised_L_traj]
self.set_block_state(state, block_state)
return components, state
|