rocm: fix hipBLASLt BGRADB wgrad failure on gfx950

miles/backends/megatron_utils/model_provider.py

@@ -93,6 +93,16 @@ def _apply_bridge_runtime_config(provider, args: argparse.Namespace) -> None:
 
     if hasattr(provider, "dsa_attention_backend"):
         provider.dsa_attention_backend = getattr(args, "dsa_attention_backend", "megatron")
+    # The bridge provider defaults gradient_accumulation_fusion=True
+    # (via fusions.can_enable_gradient_accumulation_fusion). On ROCm/gfx950,
+    # this makes TE's LayerNormLinear backward issue a bias-fused wgrad GEMM
+    # with bf16 inputs → fp32 output + HIPBLASLT_EPILOGUE_BGRADB + accumulate,
+    # for which hipBLASLt has no algorithm. Honor the Megatron CLI flag so
+    # that --no-gradient-accumulation-fusion actually takes effect.
+    if hasattr(provider, "gradient_accumulation_fusion"):
+        provider.gradient_accumulation_fusion = getattr(
+            args, "gradient_accumulation_fusion", provider.gradient_accumulation_fusion
+        )
 
 
 # Adapt from https://github.com/volcengine/verl/blob/c3b20575d2bc815fcccd84bddb4c0401fc4b632b/verl/models/llama/megatron/layers/parallel_linear.py#L82

tests/ci/rocm_utils.py

@@ -0,0 +1,5 @@
+"""Shared ROCm detection for CI tests."""
+
+import torch
+
+IS_ROCM = getattr(torch.version, "hip", None) is not None