diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index 08bedae6b0..f1a1d65c44 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -1178,6 +1178,23 @@ def run_training_process(
         except Exception:
             pass
 
+    # ── 1f. Disable torch.compile on Windows ROCm ──
+    # torch._grouped_mm crashes on gfx1200 (null HIP kernel pointer, 0xC0000005).
+    # The crash is triggered via torch.compile's JitDecomp dispatch during the
+    # first forward pass (stack: _grouped_mm ← JitDecompRegisterer ← Python).
+    # Disabling dynamo entirely avoids the kernel dispatch.  torch is already
+    # in sys.modules from section 1e's `import torch.distributed`.
+    if sys.platform == "win32" and "TORCHDYNAMO_DISABLE" not in os.environ:
+        _torch_for_rocm_check = sys.modules.get("torch")
+        if _torch_for_rocm_check is not None and getattr(
+            getattr(_torch_for_rocm_check, "version", None), "hip", None
+        ):
+            os.environ["TORCHDYNAMO_DISABLE"] = "1"
+            logger.info(
+                "Windows ROCm detected — torch.compile disabled "
+                "(_grouped_mm kernel crashes on gfx1200 with 0xC0000005)"
+            )
+
     # ── 2. Now import ML libraries (fresh in this clean process) ──
     try:
         _send_status(event_queue, "Importing Unsloth...")
diff --git a/studio/install_python_stack.py b/studio/install_python_stack.py
index ec2c4e14fe..f5b7b3c1c9 100644
--- a/studio/install_python_stack.py
+++ b/studio/install_python_stack.py
@@ -352,6 +352,19 @@ def _ensure_rocm_torch() -> None:
     # setup.ps1 sets this when it already installed AMD wheels; skip the probe.
     if os.environ.get("UNSLOTH_ROCM_TORCH_INSTALLED") == "1":
         _rocm_windows_torch_installed = True
+        # setup.ps1 already installed ROCm torch, but we still need to install
+        # the AMD Windows BNB wheel here — the PyPI bitsandbytes wheel ships
+        # only CUDA DLLs and will fail to load on ROCm (no libbitsandbytes_rocm72.dll).
+        _bnb_win_url = _BNB_ROCM_PRERELEASE_URLS.get("win_amd64")
+        if _bnb_win_url is not None:
+            pip_install_try(
+                "bitsandbytes (AMD Windows, pre-release main)",
+                "--force-reinstall",
+                "--no-cache-dir",
+                "--no-deps",
+                _bnb_win_url,
+                constrain = False,
+            )
         return
     if IS_MACOS:
         return