ci(rocm): register ROCm CI for all 8-GPU stage-c-8-gpu-mi300x tests

Add register_rocm_ci() alongside each existing register_cuda_ci() in
the stage-c-8-gpu-h100 suite so the MI300X dispatch workflow runs the
full 8-GPU test set via HIP/CUDA compatibility.

Also renames the 8-GPU ROCm suite from "stage-c-8-gpu-mi350" to
"stage-c-8-gpu-mi300x" (matching the PR stage/runner naming) in
run_suite.py and all register_rocm_ci calls.

34 test files, matching est_time + labels + disabled flags.

Author: indianspeedster

tests/ci/run_suite.py

@@ -38,7 +38,7 @@
         "stage-c-2-gpu-h200",
     ],
     HWBackend.ROCM: [
-        "stage-c-8-gpu-mi350",
+        "stage-c-8-gpu-mi300x",
         "stage-c-4-gpu-mi350",
         "stage-c-2-gpu-mi350",
     ],

tests/e2e/ckpt/test_glm47_flash_ckpt.py

@@ -1,11 +1,12 @@
 import os
 
-from tests.ci.ci_register import register_cuda_ci
+from tests.ci.ci_register import register_cuda_ci, register_rocm_ci
 
 import miles.utils.external_utils.command_utils as U
 
 # FIXME: need to modify megatron, better fix later.
 register_cuda_ci(est_time=2400, suite="stage-c-8-gpu-h100", labels=["ckpt"], disabled="Disabled due to bugs.")
+register_rocm_ci(est_time=2400, suite="stage-c-8-gpu-mi300x", labels=["ckpt"], disabled="Disabled due to bugs.")
 
 ENABLE_EVAL = 0
 USE_DEEPEP = 0

tests/e2e/ckpt/test_qwen3_4B_ckpt.py

@@ -5,7 +5,7 @@
 import miles.utils.external_utils.command_utils as U
 
 register_cuda_ci(est_time=1200, suite="stage-c-8-gpu-h100", labels=["ckpt"])
-register_rocm_ci(est_time=1200, suite="stage-c-8-gpu-mi350", labels=["ckpt"])
+register_rocm_ci(est_time=1200, suite="stage-c-8-gpu-mi300x", labels=["ckpt"])
 
 ENABLE_EVAL = bool(int(os.environ.get("MILES_TEST_ENABLE_EVAL", "1")))
 

tests/e2e/fsdp/test_qwen3_0.6B_fsdp_distributed.py

@@ -1,6 +1,6 @@
 import os
 
-from tests.ci.ci_register import register_cuda_ci
+from tests.ci.ci_register import register_cuda_ci, register_rocm_ci
 
 import miles.utils.external_utils.command_utils as U
 
@@ -10,6 +10,12 @@
     labels=["fsdp"],
     disabled="FSDP backend has known issues, not actively maintained",
 )
+register_rocm_ci(
+    est_time=600,
+    suite="stage-c-8-gpu-mi300x",
+    labels=["fsdp"],
+    disabled="FSDP backend has known issues, not actively maintained",
+)
 
 MODEL_NAME = "Qwen3-0.6B"
 NUM_GPUS = 8

tests/e2e/fsdp/test_qwen3_0.6B_megatron_fsdp_align.py

@@ -1,6 +1,6 @@
 import os
 
-from tests.ci.ci_register import register_cuda_ci
+from tests.ci.ci_register import register_cuda_ci, register_rocm_ci
 
 import miles.utils.external_utils.command_utils as U
 
@@ -10,6 +10,12 @@
     labels=["fsdp"],
     disabled="FSDP backend has known issues, not actively maintained",
 )
+register_rocm_ci(
+    est_time=900,
+    suite="stage-c-8-gpu-mi300x",
+    labels=["fsdp"],
+    disabled="FSDP backend has known issues, not actively maintained",
+)
 
 
 MODEL_NAME = "Qwen3-0.6B"

tests/e2e/fsdp/test_qwen3_4B_fsdp_true_on_policy.py

@@ -1,6 +1,6 @@
 import os
 
-from tests.ci.ci_register import register_cuda_ci
+from tests.ci.ci_register import register_cuda_ci, register_rocm_ci
 
 import miles.utils.external_utils.command_utils as U
 
@@ -10,6 +10,12 @@
     labels=["fsdp"],
     disabled="FSDP backend has known issues, not actively maintained",
 )
+register_rocm_ci(
+    est_time=600,
+    suite="stage-c-8-gpu-mi300x",
+    labels=["fsdp"],
+    disabled="FSDP backend has known issues, not actively maintained",
+)
 
 ENABLE_EVAL = bool(int(os.environ.get("MILES_TEST_ENABLE_EVAL", "1")))
 NUM_GPUS = 8

tests/e2e/megatron/test_deepseek_v32_5layer_fp8.py

@@ -13,9 +13,10 @@
     _prepare_download,
     _prepare_megatron_ckpt,
 )
-from tests.ci.ci_register import register_cuda_ci
+from tests.ci.ci_register import register_cuda_ci, register_rocm_ci
 
 register_cuda_ci(est_time=1800, suite="stage-c-8-gpu-h100", labels=["model-scripts"])
+register_rocm_ci(est_time=1800, suite="stage-c-8-gpu-mi300x", labels=["model-scripts"])
 
 
 def _args() -> ScriptArgs:

tests/e2e/megatron/test_deepseek_v4_flash_4layer_ci.py

@@ -1,9 +1,10 @@
 import os
 
 from scripts.run_deepseek_v4 import ScriptArgs, _prepare_download, _prepare_single, _prepare_spmd, _train
-from tests.ci.ci_register import register_cuda_ci
+from tests.ci.ci_register import register_cuda_ci, register_rocm_ci
 
 register_cuda_ci(est_time=1800, suite="stage-c-8-gpu-h100", labels=["model-scripts"])
+register_rocm_ci(est_time=1800, suite="stage-c-8-gpu-mi300x", labels=["model-scripts"])
 
 
 def _args() -> ScriptArgs:

tests/e2e/megatron/test_glm47_flash/test_r3_mtp.py

@@ -1,9 +1,10 @@
 import os
 
-from tests.ci.ci_register import register_cuda_ci
+from tests.ci.ci_register import register_cuda_ci, register_rocm_ci
 from tests.e2e.megatron.test_glm47_flash._common import CaseConfig, execute, prepare
 
 register_cuda_ci(est_time=1100, suite="stage-c-8-gpu-h100", labels=["megatron"])
+register_rocm_ci(est_time=1100, suite="stage-c-8-gpu-mi300x", labels=["megatron"])
 
 CASE = CaseConfig(
     use_deepep=False,

tests/e2e/megatron/test_glm47_flash/test_r3_mtp_deepep.py

@@ -1,6 +1,6 @@
 import os
 
-from tests.ci.ci_register import register_cuda_ci
+from tests.ci.ci_register import register_cuda_ci, register_rocm_ci
 from tests.e2e.megatron.test_glm47_flash._common import CaseConfig, execute, prepare
 
 # FIXME: sglang deepep code path bug.
@@ -10,6 +10,12 @@
     labels=["megatron"],
     disabled="Disabled due to sglang deepep code path bug.",
 )
+register_rocm_ci(
+    est_time=900,
+    suite="stage-c-8-gpu-mi300x",
+    labels=["megatron"],
+    disabled="Disabled due to sglang deepep code path bug.",
+)
 
 CASE = CaseConfig(
     use_deepep=True,