From 0e9387c395a71117ce10d7a5c2c80caae52f85e8 Mon Sep 17 00:00:00 2001
From: Dan Saunders <danjsaund@gmail.com>
Date: Wed, 17 Sep 2025 14:35:36 -0400
Subject: [PATCH] fix

---
 src/axolotl/kernels/moe/torch_grouped.py | 29 ++++++++++++++++++++++--
 1 file changed, 27 insertions(+), 2 deletions(-)

diff --git a/src/axolotl/kernels/moe/torch_grouped.py b/src/axolotl/kernels/moe/torch_grouped.py
index 9453ad437..18b8dcc06 100644
--- a/src/axolotl/kernels/moe/torch_grouped.py
+++ b/src/axolotl/kernels/moe/torch_grouped.py
@@ -258,7 +258,28 @@ def moe_ffn_forward_grouped(
         out = torch.zeros_like(x)
         return out.view(bsz, seqlen, hdim), router_logits
 
-    Y_list = _call_grouped_mm(As, Bs)
+    comp_dtype = dt
+    if dt == torch.bfloat16:
+        comp_dtype = torch.float16
+        if not getattr(experts_module, "_ax_grouped_logged_cast", False):
+            _LOGGER.info("torch_grouped: casting grouped_mm operands to float16")
+            experts_module._ax_grouped_logged_cast = True
+
+    def _maybe_cast(
+        tensors: List[torch.Tensor], *, to_dtype: torch.dtype
+    ) -> List[torch.Tensor]:
+        if to_dtype == dt:
+            return tensors
+        return [t.to(to_dtype) for t in tensors]
+
+    def _restore_dtype(tensors: List[torch.Tensor]) -> List[torch.Tensor]:
+        if comp_dtype == dt:
+            return tensors
+        return [t.to(dt) for t in tensors]
+
+    As_mm = _maybe_cast(As, to_dtype=comp_dtype)
+    Bs_mm = _maybe_cast(Bs, to_dtype=comp_dtype)
+    Y_list = _call_grouped_mm(As_mm, Bs_mm)
     if Y_list is None:
         if not getattr(experts_module, "_ax_grouped_logged_fail", False):
             _LOGGER.warning(
@@ -266,6 +287,7 @@ def moe_ffn_forward_grouped(
             )
             experts_module._ax_grouped_logged_fail = True
         return None, None
+    Y_list = _restore_dtype(Y_list)
 
     As2: List[torch.Tensor] = []
     Bs2: List[torch.Tensor] = []
@@ -276,7 +298,9 @@ def moe_ffn_forward_grouped(
         As2.append(Yi_hidden)
         Bs2.append(W2[i])
 
-    Y2_list = _call_grouped_mm(As2, Bs2)
+    As2_mm = _maybe_cast(As2, to_dtype=comp_dtype)
+    Bs2_mm = _maybe_cast(Bs2, to_dtype=comp_dtype)
+    Y2_list = _call_grouped_mm(As2_mm, Bs2_mm)
     if Y2_list is None:
         if not getattr(experts_module, "_ax_grouped_logged_fail", False):
             _LOGGER.warning(
@@ -284,6 +308,7 @@ def moe_ffn_forward_grouped(
             )
             experts_module._ax_grouped_logged_fail = True
         return None, None
+    Y2_list = _restore_dtype(Y2_list)
 
     for (_i, sel), Out_i in zip(expert_slices, Y2_list, strict=False):
         y_buf[sel] = Out_i