diff --git a/miles/backends/megatron_utils/model_provider.py b/miles/backends/megatron_utils/model_provider.py
index e7c24a3191..d99673f748 100644
--- a/miles/backends/megatron_utils/model_provider.py
+++ b/miles/backends/megatron_utils/model_provider.py
@@ -110,6 +110,15 @@ def wrapped_model_provider(
             provider.moe_router_bias_update_rate = args.moe_router_bias_update_rate
         if getattr(args, "moe_aux_loss_coeff", None) is not None:
             provider.moe_aux_loss_coeff = args.moe_aux_loss_coeff
+        # 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.
+        provider.gradient_accumulation_fusion = getattr(
+            args, "gradient_accumulation_fusion", provider.gradient_accumulation_fusion
+        )
         provider.finalize()
 
         def wrapped_bridge_provider(
diff --git a/tests/e2e/sglang_config/test_sglang_config_mixed_offload.py b/tests/e2e/sglang_config/test_sglang_config_mixed_offload.py
index 79c54a42a6..99d0140403 100644
--- a/tests/e2e/sglang_config/test_sglang_config_mixed_offload.py
+++ b/tests/e2e/sglang_config/test_sglang_config_mixed_offload.py
@@ -16,10 +16,13 @@
 import os
 import tempfile
 
+import torch
 from tests.ci.ci_register import register_cuda_ci
 
 import miles.utils.external_utils.command_utils as U
 
+IS_ROCM = torch.version.hip is not None
+
 register_cuda_ci(est_time=600, suite="stage-b-short-8-gpu", num_gpus=8)
 
 TIGHT_DEVICE_MEMORY = U.get_bool_env_var("MILES_TEST_TIGHT_DEVICE_MEMORY", "1")
@@ -119,13 +122,17 @@ def execute():
         f"--sglang-config {config_path} "
     )
 
-    ci_args = "--ci-test "
+    ci_args = (
+        "--ci-test "
+        + ("--ci-disable-kl-checker --ci-disable-logprobs-checker " if IS_ROCM else "")
+    )
 
     misc_args = (
         "--attention-dropout 0.0 "
         "--hidden-dropout 0.0 "
         "--accumulate-allreduce-grads-in-fp32 "
-        "--attention-softmax-in-fp32 "
+        + ("--no-gradient-accumulation-fusion " if IS_ROCM else "")
+        + "--attention-softmax-in-fp32 "
         "--attention-backend flash "
         "--actor-num-nodes 1 "
         "--actor-num-gpus-per-node 8 "