Added support for MPS (Apple Silicon)

README.md

@@ -25,6 +25,22 @@ pip install "rc-foundry[all]"
 > [!NOTE]
 > Use `pip` (not `uv`) for XPU installs since UV re-resolves dependencies and may replace your XPU torch with the standard PyPI version.
 
+**macOS (Apple Silicon) Installation**
+
+MPS support is available via a community fork. Install PyTorch first, then install directly from the fork:
+```bash
+pip install torch
+pip install "rc-foundry[all] @ git+https://github.com/fnachon/foundry.git"
+```
+
+All three models — **RFD3**, **RF3**, and **ProteinMPNN/LigandMPNN** — run on Apple Silicon MPS.
+
+> [!NOTE]
+> - The `rf3` extra (cuEquivariance) is Linux-only and is automatically skipped on macOS.
+> - Use `float32` precision — `bfloat16` is not supported on MPS. The MPS accelerator is selected and float32 precision is enforced automatically.
+> - Inference only; multi-GPU training is not supported on MPS.
+> - For `rf3 fold`, pass an absolute path to the input CIF file.
+
 **Downloading weights** Models can be downloaded to a target folder with:
 ```
 foundry install base-models --checkpoint-dir <path/to/ckpt/dir>

models/mpnn/src/mpnn/inference_engines/mpnn.py

@@ -74,6 +74,8 @@ def __init__(
             self.device = torch.device("cuda")
         elif hasattr(torch, "xpu") and torch.xpu.is_available():
             self.device = torch.device("xpu")
+        elif torch.backends.mps.is_available():
+            self.device = torch.device("mps")
         else:
             self.device = torch.device("cpu")
 
@@ -258,6 +260,8 @@ def run(
                 np.random.seed(seed)
                 if torch.cuda.is_available():
                     torch.cuda.manual_seed_all(seed)
+                elif torch.backends.mps.is_available():
+                    torch.mps.manual_seed(seed)
 
             # Run the batches for this input.
             for batch_idx in range(inference_input.input_dict["number_of_batches"]):

models/mpnn/src/mpnn/model/mpnn.py

@@ -1364,7 +1364,7 @@ def decode_auto_regressive(
         # precision settings. This works because the W_out layer is a linear
         # layer, which has predictable dtype behavior with AMP.
         device = input_features["residue_mask"].device
-        if device.type in ("cuda", "cpu") and torch.is_autocast_enabled(
+        if device.type in ("cuda", "cpu", "mps") and torch.is_autocast_enabled(
             device_type=device.type
         ):
             output_dtype = torch.get_autocast_dtype(device_type=device.type)

models/rf3/src/rf3/model/RF3.py

@@ -16,6 +16,7 @@
 from torch import nn
 
 from foundry.training.checkpoint import create_custom_forward
+from foundry.utils.torch import device_of
 
 """
 Shape Annotation Glossary:
@@ -148,7 +149,7 @@ def forward(
         """
         # Cast features to lower precision if autocast is enabled
         if torch.is_autocast_enabled():
-            autocast_dtype = torch.get_autocast_dtype("cuda")
+            autocast_dtype = torch.get_autocast_dtype(device_of(self).type)
             for x in [
                 "msa_stack",
                 "profile",
@@ -382,7 +383,7 @@ def forward(
         """
         # Cast features to lower precision if autocast is enabled
         if torch.is_autocast_enabled():
-            autocast_dtype = torch.get_autocast_dtype("cuda")
+            autocast_dtype = torch.get_autocast_dtype(device_of(self).type)
             for x in [
                 "msa_stack",
                 "profile",

models/rf3/src/rf3/model/layers/af3_diffusion_transformer.py

@@ -12,7 +12,7 @@
 from rf3.model.layers.mlff import ConformerEmbeddingWeightedAverage
 
 from foundry.training.checkpoint import activation_checkpointing
-from foundry.utils.torch import device_of
+from foundry.utils.torch import device_of, scatter_mean
 
 
 class AtomAttentionEncoderDiffusion(nn.Module):
@@ -241,16 +241,11 @@ def embed_atom_feats(R_L, C_L, D_LL, V_LL, P_LL, tok_idx):
             # Ensure dtype consistency for index_reduce
             processed_Q_L = processed_Q_L.to(Q_L.dtype)
 
-            A_I = (
-                torch.zeros(A_I_shape, device=Q_L.device, dtype=Q_L.dtype)
-                .index_reduce(
-                    -2,
-                    f["atom_to_token_map"].long(),
-                    processed_Q_L,
-                    "mean",
-                    include_self=False,
-                )
-                .clone()
+            A_I = scatter_mean(
+                torch.zeros(A_I_shape, device=Q_L.device, dtype=Q_L.dtype),
+                -2,
+                f["atom_to_token_map"].long(),
+                processed_Q_L,
             )
 
             return A_I, Q_L, C_L, P_LL
@@ -427,7 +422,7 @@ def forward(
             # zero out layer norms for the key and query
             return self.atom_attention(A_I, S_I, Z_II)
 
-        if self.use_deepspeed_evo or self.force_bfloat16:
+        if (self.use_deepspeed_evo or self.force_bfloat16) and A_I.device.type != "mps":
             A_I = A_I.to(torch.bfloat16)
             assert len(A_I.shape) == 3, f"(Diffusion batch, I, C_a) but got {A_I.shape}"
 

models/rf3/src/rf3/model/layers/attention.py

@@ -101,7 +101,7 @@ def _forward_cuequivariance(self, pair, bias):
         """cuEquivariance triangle attention implementation."""
         # Handle autocast conversion
         if torch.is_autocast_enabled():
-            dtype = torch.get_autocast_dtype("cuda")
+            dtype = torch.get_autocast_dtype(pair.device.type)
             pair = pair.to(dtype=dtype)
             bias = bias.to(dtype=dtype)
 
@@ -288,7 +288,7 @@ def _forward_cuequivariance(
         # Handle autocast conversion
         # (Use bfloat16 for optimal performance)
         if torch.is_autocast_enabled():
-            dtype = torch.get_autocast_dtype("cuda")
+            dtype = torch.get_autocast_dtype(pair.device.type)
             pair = pair.to(dtype=dtype)
 
         assert (

models/rf3/src/rf3/model/layers/pairformer_layers.py

@@ -21,6 +21,7 @@
 
 from foundry.model.layers.blocks import Dropout
 from foundry.training.checkpoint import activation_checkpointing
+from foundry.utils.torch import scatter_mean
 
 
 class AtomAttentionEncoderPairformer(nn.Module):
@@ -198,17 +199,12 @@ def embed_features(C_L, D_LL, V_LL):
             # Ensure dtype consistency for index_reduce
             processed_Q_L = processed_Q_L.to(Q_L.dtype)
 
-            A_I = torch.zeros(
-                A_I_shape, device=Q_L.device, dtype=Q_L.dtype
-            ).index_reduce(
-                -2,  # Operate on the second-to-last dimension (the atom dimension)
-                f[
-                    "atom_to_token_map"
-                ].long(),  # [L], mapping from atom index to token index. Must be a torch.int64 or torch.int32 tensor.
-                processed_Q_L,  # [L, C_atom] -> [L, C_token]
-                "mean",
-                include_self=False,  # Do not use the original values in A_I (all zeros) when aggregating
-            )  # [L, C_atom] -> [I, C_token]
+            A_I = scatter_mean(
+                torch.zeros(A_I_shape, device=Q_L.device, dtype=Q_L.dtype),
+                -2,
+                f["atom_to_token_map"].long(),  # [L], mapping from atom index to token index
+                processed_Q_L,  # (..., L, C_token)
+            )  # (..., I, C_token)
 
             return A_I, Q_L, C_L, P_LL
 
@@ -253,7 +249,7 @@ def forward(
         assert S_I is None
         A_I = self.ln_1(A_I)
 
-        if self.use_deepspeed_evo or self.force_bfloat16:
+        if (self.use_deepspeed_evo or self.force_bfloat16) and A_I.device.type != "mps":
             A_I = A_I.to(torch.bfloat16)
 
         Q_IH = self.to_q(A_I)  # / np.sqrt(self.c)
@@ -266,7 +262,7 @@ def forward(
 
         if not self.use_deepspeed_evo or L <= 24:
             Q_IH = Q_IH / torch.sqrt(
-                torch.tensor(self.c).to(Q_IH.device, torch.bfloat16)
+                torch.tensor(self.c).to(Q_IH.device, Q_IH.dtype)
             )
             # Attention
             A_IIH = torch.softmax(

models/rfd3/src/rfd3/inference/symmetry/frames.py

@@ -178,7 +178,8 @@ def _mean_along_dim(X, dim):
     R = U @ V
     if is_torch:
         F = torch.eye(3, 3, device=R.device).expand(B, 3, 3).clone()
-        F[..., -1, -1] = torch.sign(torch.linalg.det(R))
+        det = torch.linalg.det(R)
+        F[..., -1, -1] = torch.sign(det)
     else:
         F = np.broadcast_to(np.eye(3, 3), (B, 3, 3)).copy()
         F[..., -1, -1] = np.sign(np.linalg.det(R))

models/rfd3/src/rfd3/model/RFD3_diffusion_module.py

@@ -239,8 +239,6 @@ def forward(
             Q_L = self.encoder(Q_L, C_L, P_LL, indices=f["attn_indices"])
         A_I = self.downcast_q(Q_L, A_I=A_I, S_I=S_I, tok_idx=tok_idx)
 
-        # Debug chunked parameters
-
         # ... Run forward with recycling
         recycled_features = self.forward_with_recycle(
             n_recycle,
@@ -340,7 +338,6 @@ def process_(
             ),
             full=not (os.environ.get("RFD3_LOW_MEMORY_MODE", None) == "1"),
         )
-
         # ... Decoder readout
         # Check if using chunked P_LL mode
 

models/rfd3/src/rfd3/model/layers/block_utils.py

@@ -54,6 +54,11 @@ def build_valid_mask(
     return valid_mask
 
 
+def _atom_flat_idx(valid_mask: torch.Tensor) -> torch.Tensor:
+    """Return the 1-D indices of valid atoms in the flattened (n_tokens * A) grid."""
+    return valid_mask.flatten().nonzero(as_tuple=False).squeeze(1)
+
+
 def ungroup_atoms(Q_L, valid_mask):
     """
     Args
@@ -67,11 +72,20 @@ def ungroup_atoms(Q_L, valid_mask):
     """
     B, n_atoms, c = Q_L.shape
     n_tokens, A = valid_mask.shape
-    Q_IA = torch.zeros(B, n_tokens, A, c, dtype=Q_L.dtype, device=Q_L.device)
-    mask4d = valid_mask.unsqueeze(0).unsqueeze(-1)  # (1, n_tok, A, 1)
-    mask4d = mask4d.expand(B, -1, -1, c)  # (B, n_tok, A, c)
-    Q_IA.masked_scatter_(mask4d, Q_L)
-    return Q_IA
+    if Q_L.device.type == "mps":
+        # masked_scatter_ with non-contiguous masks is unreliable on MPS;
+        # use scatter with integer indices instead.
+        flat_idx = _atom_flat_idx(valid_mask)  # (n_atoms,)
+        idx = flat_idx.view(1, -1, 1).expand(B, -1, c)  # (B, n_atoms, c)
+        Q_IA = torch.zeros(B, n_tokens * A, c, dtype=Q_L.dtype, device=Q_L.device)
+        Q_IA = Q_IA.scatter(1, idx, Q_L)
+        return Q_IA.reshape(B, n_tokens, A, c)
+    else:
+        Q_IA = torch.zeros(B, n_tokens, A, c, dtype=Q_L.dtype, device=Q_L.device)
+        mask4d = valid_mask.unsqueeze(0).unsqueeze(-1)  # (1, n_tok, A, 1)
+        mask4d = mask4d.expand(B, -1, -1, c)  # (B, n_tok, A, c)
+        Q_IA.masked_scatter_(mask4d, Q_L)
+        return Q_IA
 
 
 def group_atoms(Q_IA: torch.Tensor, valid_mask: torch.Tensor) -> torch.Tensor:
@@ -85,10 +99,17 @@ def group_atoms(Q_IA: torch.Tensor, valid_mask: torch.Tensor) -> torch.Tensor:
     -------
     Q_L        : (B, n_atoms, c)  flattened real atoms, order preserved
     """
-    B, _, _, c = Q_IA.shape
-    mask4d = valid_mask.unsqueeze(0).unsqueeze(-1).expand(B, -1, -1, c)  # (B,n_tok,A,c)
-    Q_L = Q_IA[mask4d].view(B, -1, c)  # restore 2‑D shape
-    return Q_L
+    B, n_tok, A, c = Q_IA.shape
+    if Q_IA.device.type == "mps":
+        # Boolean indexing with non-contiguous expanded masks is unreliable on MPS;
+        # use integer index gather instead.
+        flat_idx = _atom_flat_idx(valid_mask)  # (n_atoms,)
+        Q_L = Q_IA.reshape(B, n_tok * A, c)[:, flat_idx, :]
+        return Q_L.contiguous()
+    else:
+        mask4d = valid_mask.unsqueeze(0).unsqueeze(-1).expand(B, -1, -1, c)
+        Q_L = Q_IA[mask4d].view(B, -1, c)
+        return Q_L
 
 
 def group_pair(P_IAA, valid_mask):
@@ -137,9 +158,9 @@ def scatter_add_pair_features(P_LK_tgt, P_LK_indices, P_LA_src, P_LA_indices):
     # Handle case when indices and P_LA don't have batch dimensions
     B, L, k = P_LK_indices.shape
     if P_LA_indices.ndim == 2:
-        P_LA_indices = P_LA_indices.unsqueeze(0).expand(B, -1, -1)
+        P_LA_indices = P_LA_indices.unsqueeze(0).expand(B, -1, -1).contiguous()
     if P_LA_src.ndim == 3:
-        P_LA_src = P_LA_src.unsqueeze(0).expand(B, -1, -1)
+        P_LA_src = P_LA_src.unsqueeze(0).expand(B, -1, -1).contiguous()
     assert (
         P_LA_src.shape[-1] == P_LK_tgt.shape[-1]
     ), "Channel dims do not match, got: {} vs {}".format(
@@ -154,8 +175,8 @@ def scatter_add_pair_features(P_LK_tgt, P_LK_indices, P_LA_src, P_LA_indices):
     k_indices = matches.long().argmax(dim=-1)  # (B, L, a)
     scatter_indices = k_indices.unsqueeze(-1).expand(
         -1, -1, -1, P_LK_tgt.shape[-1]
-    )  # (B, L, a, c)
-    P_LK_tgt = P_LK_tgt.scatter_add(dim=2, index=scatter_indices, src=P_LA_src)
+    ).contiguous()  # (B, L, a, c)
+    P_LK_tgt = P_LK_tgt.scatter_add(dim=2, index=scatter_indices, src=P_LA_src.contiguous())
     return P_LK_tgt
 
 
@@ -169,8 +190,8 @@ def _batched_gather(values: torch.Tensor, idx: torch.Tensor) -> torch.Tensor:
     k = idx.shape[-1]
 
     #   (B, L, 1, C)  → stride-0 along k  → (B, L, k, C)
-    src = values.unsqueeze(2).expand(-1, -1, k, -1)
-    idx = idx.unsqueeze(-1).expand(-1, -1, -1, C)  # (B, L, k, C)
+    src = values.unsqueeze(2).expand(-1, -1, k, -1).contiguous()
+    idx = idx.unsqueeze(-1).expand(-1, -1, -1, C).contiguous()  # (B, L, k, C)
 
     return torch.gather(src, 1, idx)  # dim=1 is the L-axis
 
@@ -350,7 +371,7 @@ def build_index_mask(
     # Exclude tokens which are partially filled (L, I)
     n_query_per_token = torch.zeros((L, I), device=device).float()
     n_query_per_token.scatter_add_(
-        1, tok_idx.long()[None, :].expand(L, -1), mask.float()
+        1, tok_idx.long()[None, :].expand(L, -1).contiguous(), mask.float()
     )
 
     # Find mask for the atoms for which the number of keys
@@ -407,21 +428,23 @@ def extend_index_mask_with_neighbours(
     inf = torch.tensor(float("inf"), dtype=D_LL.dtype, device=device)
 
     # 1. Selection of sequence neighbours
-    all_idx_row = torch.arange(L, device=device).expand(L, L)
-    indices = torch.where(mask, all_idx_row, inf)  # sentinel inf if not-forced
+    # Use .repeat() instead of .expand() to produce a contiguous tensor — MPS does
+    # not handle non-contiguous inputs to torch.where correctly.
+    all_idx_row = torch.arange(L, device=device).unsqueeze(0).repeat(L, 1)
+    indices = torch.where(mask.contiguous(), all_idx_row, inf)  # sentinel inf if not-forced
     indices = indices.sort(dim=1)[0][:, :k]  # (L, k)
 
     # 2. Find k-nn excluding forced indices
-    D_LL = torch.where(mask, inf, D_LL)
+    D_LL = torch.where(mask.contiguous(), inf, D_LL)
     filler_idx = torch.topk(D_LL, k, dim=-1, largest=False).indices
 
     # ... Reverse last axis s.t. best matched indices are last
     filler_idx = filler_idx.flip(dims=[-1])
 
     # 3. Fill indices
     to_fill = indices == inf
-    to_fill = to_fill.expand_as(filler_idx)
-    indices = indices.expand_as(filler_idx)
+    to_fill = to_fill.expand_as(filler_idx).contiguous()
+    indices = indices.expand_as(filler_idx).contiguous()
     indices = torch.where(to_fill, filler_idx, indices)
 
     return indices.long()  # (B, L, k)
@@ -437,7 +460,7 @@ def get_sparse_attention_indices(
 
     # Sort and assert no duplicates (optional but good practise)
     indices, _ = torch.sort(indices, dim=-1)
-    if (indices[..., 1:] == indices[..., :-1]).any():
+    if indices.device.type != "mps" and (indices[..., 1:] == indices[..., :-1]).any():
         raise AssertionError("Tensor has duplicate elements along the last dimension.")
 
     assert (

models/rfd3/src/rfd3/model/layers/blocks.py

@@ -30,6 +30,7 @@
 
 from foundry import DISABLE_CHECKPOINTING
 from foundry.common import exists
+from foundry.utils.torch import scatter_mean
 
 logger = logging.getLogger(__name__)
 
@@ -213,16 +214,11 @@ def forward(self, R_L, tok_idx):
             self.c_token,
         )
         Q_L = self.linear(R_L)
-        A_I = (
-            torch.zeros(A_I_shape, device=R_L.device, dtype=Q_L.dtype)
-            .index_reduce(
-                -2,
-                tok_idx.long(),
-                Q_L,
-                "mean",
-                include_self=False,
-            )
-            .clone()
+        A_I = scatter_mean(
+            torch.zeros(A_I_shape, device=R_L.device, dtype=Q_L.dtype),
+            -2,
+            tok_idx.long(),
+            Q_L,
         )
         return A_I