diff --git a/buildkite/pipeline_generator/buildkite_step.py b/buildkite/pipeline_generator/buildkite_step.py
index 42c1dd6b..82881965 100644
--- a/buildkite/pipeline_generator/buildkite_step.py
+++ b/buildkite/pipeline_generator/buildkite_step.py
@@ -299,8 +299,12 @@ def _create_amd_mirror_step(step: Step, original_commands: List[str], amd: Dict[
     if working_dir:
         amd_commands_str = f"cd {working_dir} && {amd_commands_str}"
 
-    # Add AMD test script wrapper
-    amd_command_wrapped = f'bash .buildkite/scripts/hardware_ci/run-amd-test.sh "{amd_commands_str}"'
+    # Pass commands via VLLM_TEST_COMMANDS env var instead of positional
+    # argument. This preserves all inner quoting (double quotes from
+    # _prepare_commands' single-to-double conversion, pytest -m/-k
+    # expressions, etc.) because Buildkite sets env vars directly in the
+    # process environment without shell interpretation.
+    amd_command_wrapped = "bash .buildkite/scripts/hardware_ci/run-amd-test.sh"
 
     # Extract device name from queue name
     device_type = amd_device.replace("amd_", "") if amd_device.startswith("amd_") else amd_device
@@ -335,7 +339,7 @@ def _create_amd_mirror_step(step: Step, original_commands: List[str], amd: Dict[
         commands=[amd_command_wrapped],
         depends_on=["image-build-amd"],
         agents={"queue": amd_queue},
-        env={"DOCKER_BUILDKIT": "1"},
+        env={"DOCKER_BUILDKIT": "1", "VLLM_TEST_COMMANDS": amd_commands_str},
         priority=200,
         soft_fail=False,
         retry=amd_retry,
diff --git a/buildkite/test-template-amd.j2 b/buildkite/test-template-amd.j2
index de5bb7db..1131d532 100644
--- a/buildkite/test-template-amd.j2
+++ b/buildkite/test-template-amd.j2
@@ -356,9 +356,10 @@ plugins:
           {% else %}
           queue: amd_mi325_1
           {% endif %}
-        command: bash .buildkite/scripts/hardware_ci/run-amd-test.sh "(command rocm-smi || true) && export VLLM_TEST_GROUP_NAME={{ step.agent_pool }}-{{ step.label | replace(" ", "-") | lower | replace("(", "") | replace(")", "") | replace("%", "") | replace(",", "-") | replace("+", "-") }} && export VLLM_ALLOW_DEPRECATED_BEAM_SEARCH=1 && cd {{ (step.working_dir or default_working_dir) | safe  }} ; {{ step.command  or (step.commands | join(" && ")) | safe }}"
+        command: bash .buildkite/scripts/hardware_ci/run-amd-test.sh
         env:
           DOCKER_BUILDKIT: "1"
+          VLLM_TEST_COMMANDS: "(command rocm-smi || true) && export VLLM_TEST_GROUP_NAME={{ step.agent_pool }}-{{ step.label | replace(' ', '-') | lower | replace('(', '') | replace(')', '') | replace('%', '') | replace(',', '-') | replace('+', '-') }} && export VLLM_ALLOW_DEPRECATED_BEAM_SEARCH=1 && cd {{ (step.working_dir or default_working_dir) | safe }} && {{ step.command or (step.commands | join(' && ')) | safe }}"
         priority: 100
         {% if step.grade and step.grade == "Blocking" %}
         soft_fail: false
diff --git a/buildkite/test-template-ci.j2 b/buildkite/test-template-ci.j2
index 96638c01..5b8e6754 100644
--- a/buildkite/test-template-ci.j2
+++ b/buildkite/test-template-ci.j2
@@ -612,9 +612,10 @@ steps:
          {% else %}
            queue: amd_mi325_1
          {% endif%}
-        command: bash .buildkite/scripts/hardware_ci/run-amd-test.sh "(command rocm-smi || true) && export VLLM_ALLOW_DEPRECATED_BEAM_SEARCH=1 && cd {{ (step.working_dir or default_working_dir) | safe  }} ; {{ step.command  or (step.commands | join(" && ")) | safe }}"
+        command: bash .buildkite/scripts/hardware_ci/run-amd-test.sh
         env:
           DOCKER_BUILDKIT: "1"
+          VLLM_TEST_COMMANDS: "(command rocm-smi || true) && export VLLM_ALLOW_DEPRECATED_BEAM_SEARCH=1 && cd {{ (step.working_dir or default_working_dir) | safe }} && {{ step.command or (step.commands | join(' && ')) | safe }}"
         priority: 100
         {% endif %}
     {% endfor %}