wafer-cli 0.2.8__py3-none-any.whl → 0.2.10__py3-none-any.whl

wafer/evaluate.py

@@ -14,6 +14,7 @@ logger = logging.getLogger(__name__)
 from wafer_core.utils.kernel_utils.targets.config import (
     BaremetalTarget,
     DigitalOceanTarget,
+    LocalTarget,
     ModalTarget,
     RunPodTarget,
     VMTarget,
@@ -21,6 +22,30 @@ from wafer_core.utils.kernel_utils.targets.config import (
 )
 
 
+# Map AMD compute capability to ROCm architecture
+# Used to set PYTORCH_ROCM_ARCH for faster compilation (compile only for target arch)
+AMD_CC_TO_ARCH = {
+    "9.4": "gfx942",  # MI300X
+    "9.0a": "gfx90a",  # MI200 series
+    "9.08": "gfx908",  # MI100
+    "9.06": "gfx906",  # MI50/60
+    "10.30": "gfx1030",  # RDNA2
+    "11.0": "gfx1100",  # RDNA3
+}
+
+
+def _get_rocm_arch(compute_capability: str) -> str | None:
+    """Get ROCm architecture string from compute capability.
+
+    Returns gfx* string for PYTORCH_ROCM_ARCH, or None if not found.
+    """
+    # Already a gfx string
+    if compute_capability.startswith("gfx"):
+        return compute_capability
+    # Map from numeric CC
+    return AMD_CC_TO_ARCH.get(compute_capability)
+
+
 def _build_docker_run_command(
     image: str,
     command: str,
@@ -161,6 +186,7 @@ class KernelBenchEvaluateArgs:
     inputs: Path | None = None  # Custom inputs file to override get_inputs()
     seed: int = 42  # Random seed for reproducibility
     defensive: bool = False
+    backend: str | None = None  # Kernel backend for static validation
     sync_artifacts: bool = True
     gpu_id: int | None = None
 
@@ -396,33 +422,6 @@ async def run_evaluate_docker(
     print(f"Connecting to {target.ssh_target}...")
 
     async with AsyncSSHClient(target.ssh_target, target.ssh_key) as client:
-        # Upload wafer-core to remote
-        try:
-            wafer_root = _get_wafer_root()
-            wafer_core_path = wafer_root / "packages" / "wafer-core"
-            print(f"Uploading wafer-core from {wafer_core_path}...")
-
-            # Create workspace and upload
-            workspace_name = wafer_core_path.name
-            remote_workspace = f"{REMOTE_WORKSPACE_BASE}/{workspace_name}"
-            await client.exec(f"mkdir -p {remote_workspace}")
-            wafer_core_workspace = await client.expand_path(remote_workspace)
-
-            upload_result = await client.upload_files(
-                str(wafer_core_path), wafer_core_workspace, recursive=True
-            )
-            print(f"Uploaded {upload_result.files_copied} files")
-        except Exception as e:
-            return EvaluateResult(
-                success=False,
-                all_correct=False,
-                correctness_score=0.0,
-                geomean_speedup=0.0,
-                passed_tests=0,
-                total_tests=0,
-                error_message=f"Failed to upload wafer-core: {e}",
-            )
-
         print(f"Using Docker image: {target.docker_image}")
         print(f"Using GPU {gpu_id}...")
 
@@ -431,10 +430,13 @@ async def run_evaluate_docker(
         ref_code = args.reference.read_text()
         test_cases_data = json.loads(args.test_cases.read_text())
 
-        # Create a unique run directory
+        # Create workspace for evaluation files
         timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
         run_dir = f"wafer_eval_{timestamp}"
-        run_path = f"{wafer_core_workspace}/{run_dir}"
+        eval_workspace = f"{REMOTE_WORKSPACE_BASE}/eval_{timestamp}"
+        await client.exec(f"mkdir -p {eval_workspace}")
+        eval_workspace_expanded = await client.expand_path(eval_workspace)
+        run_path = f"{eval_workspace_expanded}/{run_dir}"
 
         print("Uploading evaluation files...")
 
@@ -521,17 +523,16 @@ async def run_evaluate_docker(
         container_impl_path = f"{container_run_path}/implementation.py"
         container_ref_path = f"{container_run_path}/reference.py"
         container_test_cases_path = f"{container_run_path}/test_cases.json"
-        container_evaluate_script = (
-            f"{CONTAINER_WORKSPACE}/wafer_core/utils/kernel_utils/evaluate.py"
-        )
 
-        # Build pip install command for torch and other deps (no wafer-core install needed)
+        # Build pip install command for torch and other deps, plus wafer-core
         pip_install_cmd = _build_docker_pip_install_cmd(target)
+        install_cmd = (
+            f"{pip_install_cmd} && uv pip install --system --break-system-packages wafer-core"
+        )
 
-        # Build evaluate command - use PYTHONPATH instead of installing wafer-core
+        # Build evaluate command using installed wafer-core module
         python_cmd_parts = [
-            f"PYTHONPATH={CONTAINER_WORKSPACE}:$PYTHONPATH",
-            f"python3 {container_evaluate_script}",
+            "python3 -m wafer_core.utils.kernel_utils.evaluate",
             f"--implementation {container_impl_path}",
             f"--reference {container_ref_path}",
             f"--test-cases {container_test_cases_path}",
@@ -547,8 +548,8 @@ async def run_evaluate_docker(
 
         eval_cmd = " ".join(python_cmd_parts)
 
-        # Full command: install torch deps, then run evaluate with PYTHONPATH
-        full_cmd = f"{pip_install_cmd} && cd {container_run_path} && {eval_cmd}"
+        # Full command: install deps + wafer-core, then run evaluate
+        full_cmd = f"{install_cmd} && cd {container_run_path} && {eval_cmd}"
 
         # Build Docker run command
         # Add SYS_ADMIN capability when profiling (needed for NCU GPU performance counters)
@@ -558,7 +559,7 @@ async def run_evaluate_docker(
             working_dir=container_run_path,
             env={"CUDA_VISIBLE_DEVICES": str(gpu_id), "PYTHONUNBUFFERED": "1"},
             gpus="all",
-            volumes={wafer_core_workspace: CONTAINER_WORKSPACE},
+            volumes={eval_workspace_expanded: CONTAINER_WORKSPACE},
             cap_add=["SYS_ADMIN"] if args.profile else None,
         )
 
@@ -567,7 +568,7 @@ async def run_evaluate_docker(
         # Run Docker command and stream output
         log_lines = []
         async for line in client.exec_stream(docker_cmd):
-            print(line)
+            print(line, flush=True)
             log_lines.append(line)
 
         # Read results
@@ -665,6 +666,181 @@ async def run_evaluate_docker(
         )
 
 
+async def run_evaluate_local(
+    args: EvaluateArgs,
+    target: LocalTarget,
+) -> EvaluateResult:
+    """Run evaluation locally on the current machine.
+
+    For LocalTarget - no SSH needed, runs directly.
+
+    Args:
+        args: Evaluate arguments
+        target: Local target config
+
+    Returns:
+        Evaluation result
+    """
+    import os
+    import subprocess
+    import tempfile
+    from datetime import datetime
+
+    # Select GPU
+    gpu_id = args.gpu_id if args.gpu_id is not None else target.gpu_ids[0]
+
+    print(f"Running local evaluation on GPU {gpu_id}...")
+
+    # Create temp directory for eval files
+    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+    with tempfile.TemporaryDirectory(prefix=f"wafer_eval_{timestamp}_") as run_path:
+        run_path = Path(run_path)
+
+        # Write implementation
+        impl_path = run_path / "implementation.py"
+        impl_path.write_text(args.implementation.read_text())
+
+        # Write reference
+        ref_path = run_path / "reference.py"
+        ref_path.write_text(args.reference.read_text())
+
+        # Write custom inputs if provided
+        inputs_path = None
+        if args.inputs:
+            inputs_path = run_path / "custom_inputs.py"
+            inputs_path.write_text(args.inputs.read_text())
+
+        # Write eval script
+        eval_script_path = run_path / "kernelbench_eval.py"
+        eval_script_path.write_text(KERNELBENCH_EVAL_SCRIPT)
+
+        # Write defense module if defensive mode is enabled
+        defense_module_path = None
+        if args.defensive:
+            defense_src = (
+                Path(__file__).parent.parent.parent.parent
+                / "packages"
+                / "wafer-core"
+                / "wafer_core"
+                / "utils"
+                / "kernel_utils"
+                / "defense.py"
+            )
+            if defense_src.exists():
+                defense_module_path = run_path / "defense.py"
+                defense_module_path.write_text(defense_src.read_text())
+            else:
+                print(f"Warning: defense.py not found at {defense_src}")
+
+        # Output file
+        output_path = run_path / "results.json"
+
+        # Build eval command
+        cmd_parts = [
+            "python3",
+            str(eval_script_path),
+            "--impl",
+            str(impl_path),
+            "--reference",
+            str(ref_path),
+            "--output",
+            str(output_path),
+            "--seed",
+            str(args.seed),
+        ]
+
+        if args.benchmark:
+            cmd_parts.append("--benchmark")
+        if args.profile:
+            cmd_parts.append("--profile")
+        if inputs_path:
+            cmd_parts.extend(["--inputs", str(inputs_path)])
+        if args.defensive and defense_module_path:
+            cmd_parts.extend(["--defensive", "--defense-module", str(defense_module_path)])
+
+        # Set environment for GPU selection
+        env = os.environ.copy()
+        if target.vendor == "nvidia":
+            env["CUDA_VISIBLE_DEVICES"] = str(gpu_id)
+        else:  # AMD
+            env["HIP_VISIBLE_DEVICES"] = str(gpu_id)
+            env["ROCM_PATH"] = "/opt/rocm"
+
+        print(f"Running: {' '.join(cmd_parts[:4])} ...")
+
+        # Run evaluation
+        try:
+            result = subprocess.run(
+                cmd_parts,
+                cwd=str(run_path),
+                env=env,
+                capture_output=True,
+                text=True,
+                timeout=args.timeout or 600,
+            )
+        except subprocess.TimeoutExpired:
+            return EvaluateResult(
+                success=False,
+                all_correct=False,
+                correctness_score=0.0,
+                geomean_speedup=0.0,
+                passed_tests=0,
+                total_tests=0,
+                error_message="Evaluation timed out",
+            )
+
+        if result.returncode != 0:
+            error_msg = result.stderr or result.stdout or "Unknown error"
+            # Truncate long errors
+            if len(error_msg) > 1000:
+                error_msg = error_msg[:500] + "\n...\n" + error_msg[-500:]
+            return EvaluateResult(
+                success=False,
+                all_correct=False,
+                correctness_score=0.0,
+                geomean_speedup=0.0,
+                passed_tests=0,
+                total_tests=0,
+                error_message=f"Evaluation failed:\n{error_msg}",
+            )
+
+        # Parse results
+        if not output_path.exists():
+            return EvaluateResult(
+                success=False,
+                all_correct=False,
+                correctness_score=0.0,
+                geomean_speedup=0.0,
+                passed_tests=0,
+                total_tests=0,
+                error_message="No results.json produced",
+            )
+
+        try:
+            results = json.loads(output_path.read_text())
+        except json.JSONDecodeError as e:
+            return EvaluateResult(
+                success=False,
+                all_correct=False,
+                correctness_score=0.0,
+                geomean_speedup=0.0,
+                passed_tests=0,
+                total_tests=0,
+                error_message=f"Failed to parse results: {e}",
+            )
+
+        # Extract results
+        return EvaluateResult(
+            success=True,
+            all_correct=results.get("all_correct", False),
+            correctness_score=results.get("correctness_score", 0.0),
+            geomean_speedup=results.get("geomean_speedup", 0.0),
+            passed_tests=results.get("passed_tests", 0),
+            total_tests=results.get("total_tests", 0),
+            benchmark_results=results.get("benchmark", {}),
+        )
+
+
 async def run_evaluate_ssh(
     args: EvaluateArgs,
     target: BaremetalTarget | VMTarget,
@@ -982,6 +1158,7 @@ def _build_modal_sandbox_script(
     test_cases_b64: str,
     run_benchmarks: bool,
     run_defensive: bool,
+    defense_code_b64: str | None = None,
 ) -> str:
     """Build Python script to create sandbox and run evaluation.
 
@@ -1062,6 +1239,20 @@ print('Files written')
             print(json.dumps({{"error": f"Failed to write files: {{proc.stderr.read()}}"}}))
             return
 
+        # Write defense module if defensive mode is enabled
+        # NOTE: Check for actual base64 content, not just truthy string (None becomes "None")
+        if {run_defensive} and "{defense_code_b64}" and "{defense_code_b64}" != "None":
+            proc = sandbox.exec("python", "-c", f"""
+import base64
+with open('/workspace/defense.py', 'w') as f:
+    f.write(base64.b64decode('{defense_code_b64}').decode())
+print('Defense module written')
+""")
+            proc.wait()
+            if proc.returncode != 0:
+                print(json.dumps({{"error": f"Failed to write defense module: {{proc.stderr.read()}}"}}))
+                return
+
         # Build inline evaluation script
         eval_script = """
 import json
@@ -1089,6 +1280,26 @@ generate_input = load_fn('reference.py', 'generate_input')
 
 import torch
 
+# Load defense module if available and defensive mode is enabled
+run_defensive = {run_defensive}
+defense = None
+if run_defensive:
+    try:
+        defense = load_fn('defense.py', 'run_all_defenses')
+        time_with_defenses = load_fn('defense.py', 'time_execution_with_defenses')
+        print('[Defense] Defense module loaded')
+
+        # Wrap kernels for defense API compatibility
+        # Defense API calls kernel(*args), but functional format expects kernel(inputs_tuple)
+        # These wrappers repack the unpacked args back into a tuple
+        def _wrap_for_defense(kernel):
+            return lambda *args: kernel(args)
+        custom_kernel_for_defense = _wrap_for_defense(custom_kernel)
+        ref_kernel_for_defense = _wrap_for_defense(ref_kernel)
+    except Exception as e:
+        print(f'[Defense] Warning: Could not load defense module: {{e}}')
+        defense = None
+
 results = []
 all_correct = True
 total_time_ms = 0.0
@@ -1116,36 +1327,63 @@ for tc in test_cases:
         impl_time_ms = 0.0
         ref_time_ms = 0.0
         if {run_benchmarks}:
-            # Warmup
-            for _ in range(3):
-                custom_kernel(inputs)
-            torch.cuda.synchronize()
-
-            # Measure with defensive timing if requested
-            # Defensive: sync before recording end event to catch stream injection
-            start = torch.cuda.Event(enable_timing=True)
-            end = torch.cuda.Event(enable_timing=True)
-            start.record()
-            for _ in range(10):
-                custom_kernel(inputs)
-            if {run_defensive}:
-                torch.cuda.synchronize()  # DEFENSE: sync all streams before end
-            end.record()
-            torch.cuda.synchronize()
-            impl_time_ms = start.elapsed_time(end) / 10
-
-            # Reference timing (same defensive approach)
-            for _ in range(3):
-                ref_kernel(inputs)
-            torch.cuda.synchronize()
-            start.record()
-            for _ in range(10):
-                ref_kernel(inputs)
-            if {run_defensive}:
-                torch.cuda.synchronize()  # DEFENSE: sync all streams before end
-            end.record()
-            torch.cuda.synchronize()
-            ref_time_ms = start.elapsed_time(end) / 10
+            if run_defensive and defense is not None:
+                # Use full defense suite with wrapped kernels
+                # inputs_list unpacks the tuple so defense can infer dtype/device from tensors
+                inputs_list = list(inputs) if hasattr(inputs, '__iter__') and not isinstance(inputs, torch.Tensor) else [inputs]
+
+                # Run defense checks
+                all_passed, defense_results, _ = defense(custom_kernel_for_defense, *inputs_list)
+                if not all_passed:
+                    failed = [name for name, passed, _ in defense_results if not passed]
+                    raise ValueError(f"Defense checks failed: {{failed}}")
+
+                # Time with defensive timing (using wrapped kernels)
+                impl_times, _ = time_with_defenses(
+                    custom_kernel_for_defense,
+                    inputs_list,
+                    num_warmup=3,
+                    num_trials=10,
+                    verbose=False,
+                    run_defenses=False,
+                )
+                impl_time_ms = sum(impl_times) / len(impl_times)
+
+                ref_times, _ = time_with_defenses(
+                    ref_kernel_for_defense,
+                    inputs_list,
+                    num_warmup=3,
+                    num_trials=10,
+                    verbose=False,
+                    run_defenses=False,
+                )
+                ref_time_ms = sum(ref_times) / len(ref_times)
+            else:
+                # Standard timing without full defenses
+                # Warmup
+                for _ in range(3):
+                    custom_kernel(inputs)
+                torch.cuda.synchronize()
+
+                start = torch.cuda.Event(enable_timing=True)
+                end = torch.cuda.Event(enable_timing=True)
+                start.record()
+                for _ in range(10):
+                    custom_kernel(inputs)
+                end.record()
+                torch.cuda.synchronize()
+                impl_time_ms = start.elapsed_time(end) / 10
+
+                # Reference timing
+                for _ in range(3):
+                    ref_kernel(inputs)
+                torch.cuda.synchronize()
+                start.record()
+                for _ in range(10):
+                    ref_kernel(inputs)
+                end.record()
+                torch.cuda.synchronize()
+                ref_time_ms = start.elapsed_time(end) / 10
 
             total_time_ms += impl_time_ms
             ref_total_time_ms += ref_time_ms
@@ -1197,7 +1435,7 @@ print(json.dumps({{
         # Find the last JSON line in output
         for line in reversed(stdout.strip().split("\\n")):
             if line.startswith("{{"):
-                print(line)
+                print(line, flush=True)
                 return
 
         print(json.dumps({{"error": f"No result JSON in output: {{stdout[:500]}}"}}))
@@ -1238,6 +1476,23 @@ async def run_evaluate_modal(
     ref_code_b64 = base64.b64encode(args.reference.read_bytes()).decode()
     test_cases_b64 = base64.b64encode(args.test_cases.read_bytes()).decode()
 
+    # Encode defense module if defensive mode is enabled
+    defense_code_b64 = None
+    if args.defensive:
+        defense_path = (
+            Path(__file__).parent.parent.parent.parent
+            / "packages"
+            / "wafer-core"
+            / "wafer_core"
+            / "utils"
+            / "kernel_utils"
+            / "defense.py"
+        )
+        if defense_path.exists():
+            defense_code_b64 = base64.b64encode(defense_path.read_bytes()).decode()
+        else:
+            print(f"Warning: defense.py not found at {defense_path}, falling back to basic defense")
+
     # Build the script that creates sandbox and runs eval
     script = _build_modal_sandbox_script(
         target=target,
@@ -1246,6 +1501,7 @@ async def run_evaluate_modal(
         test_cases_b64=test_cases_b64,
         run_benchmarks=args.benchmark,
         run_defensive=args.defensive,
+        defense_code_b64=defense_code_b64,
     )
 
     def _run_subprocess() -> tuple[str, str, int]:
@@ -1343,6 +1599,7 @@ def _build_workspace_eval_script(
     test_cases_json: str,
     run_benchmarks: bool,
     run_defensive: bool = False,
+    defense_code: str | None = None,
 ) -> str:
     """Build inline evaluation script for workspace exec.
 
@@ -1353,6 +1610,7 @@ def _build_workspace_eval_script(
     impl_b64 = base64.b64encode(impl_code.encode()).decode()
     ref_b64 = base64.b64encode(ref_code.encode()).decode()
     tests_b64 = base64.b64encode(test_cases_json.encode()).decode()
+    defense_b64 = base64.b64encode(defense_code.encode()).decode() if defense_code else ""
 
     return f'''
 import base64
@@ -1372,6 +1630,15 @@ with open("/tmp/kernel.py", "w") as f:
 with open("/tmp/reference.py", "w") as f:
     f.write(ref_code)
 
+# Write defense module if available
+run_defensive = {run_defensive}
+defense_b64 = "{defense_b64}"
+# NOTE: Check defense_b64 is not empty and not the string "None" (from None formatting)
+if run_defensive and defense_b64 and defense_b64 != "None":
+    defense_code = base64.b64decode(defense_b64).decode()
+    with open("/tmp/defense.py", "w") as f:
+        f.write(defense_code)
+
 # Load kernels
 def load_fn(path, name):
     spec = importlib.util.spec_from_file_location("mod", path)
@@ -1385,6 +1652,24 @@ generate_input = load_fn("/tmp/reference.py", "generate_input")
 
 import torch
 
+# Load defense module if available
+defense = None
+if run_defensive and defense_b64 and defense_b64 != "None":
+    try:
+        defense = load_fn("/tmp/defense.py", "run_all_defenses")
+        time_with_defenses = load_fn("/tmp/defense.py", "time_execution_with_defenses")
+        print("[Defense] Defense module loaded")
+
+        # Wrap kernels for defense API compatibility
+        # Defense API calls kernel(*args), but functional format expects kernel(inputs_tuple)
+        def _wrap_for_defense(kernel):
+            return lambda *args: kernel(args)
+        custom_kernel_for_defense = _wrap_for_defense(custom_kernel)
+        ref_kernel_for_defense = _wrap_for_defense(ref_kernel)
+    except Exception as e:
+        print(f"[Defense] Warning: Could not load defense module: {{e}}")
+        defense = None
+
 results = []
 all_correct = True
 total_time_ms = 0.0
@@ -1412,36 +1697,60 @@ for tc in test_cases:
         impl_time_ms = 0.0
         ref_time_ms = 0.0
         if {run_benchmarks}:
-            # Warmup
-            for _ in range(3):
-                custom_kernel(inputs)
-            torch.cuda.synchronize()
-
-            # Measure with defensive timing if requested
-            # Defensive: sync before recording end event to catch stream injection
-            start = torch.cuda.Event(enable_timing=True)
-            end = torch.cuda.Event(enable_timing=True)
-            start.record()
-            for _ in range(10):
-                custom_kernel(inputs)
-            if {run_defensive}:
-                torch.cuda.synchronize()  # DEFENSE: sync all streams before end
-            end.record()
-            torch.cuda.synchronize()
-            impl_time_ms = start.elapsed_time(end) / 10
-
-            # Reference timing (same defensive approach)
-            for _ in range(3):
-                ref_kernel(inputs)
-            torch.cuda.synchronize()
-            start.record()
-            for _ in range(10):
-                ref_kernel(inputs)
-            if {run_defensive}:
-                torch.cuda.synchronize()  # DEFENSE: sync all streams before end
-            end.record()
-            torch.cuda.synchronize()
-            ref_time_ms = start.elapsed_time(end) / 10
+            if run_defensive and defense is not None:
+                # Use full defense suite with wrapped kernels
+                inputs_list = list(inputs) if hasattr(inputs, '__iter__') and not isinstance(inputs, torch.Tensor) else [inputs]
+
+                # Run defense checks
+                all_passed, defense_results, _ = defense(custom_kernel_for_defense, *inputs_list)
+                if not all_passed:
+                    failed = [name for name, passed, _ in defense_results if not passed]
+                    raise ValueError(f"Defense checks failed: {{failed}}")
+
+                # Time with defensive timing (using wrapped kernels)
+                impl_times, _ = time_with_defenses(
+                    custom_kernel_for_defense,
+                    inputs_list,
+                    num_warmup=3,
+                    num_trials=10,
+                    verbose=False,
+                    run_defenses=False,
+                )
+                impl_time_ms = sum(impl_times) / len(impl_times)
+
+                ref_times, _ = time_with_defenses(
+                    ref_kernel_for_defense,
+                    inputs_list,
+                    num_warmup=3,
+                    num_trials=10,
+                    verbose=False,
+                    run_defenses=False,
+                )
+                ref_time_ms = sum(ref_times) / len(ref_times)
+            else:
+                # Standard timing
+                for _ in range(3):
+                    custom_kernel(inputs)
+                torch.cuda.synchronize()
+
+                start = torch.cuda.Event(enable_timing=True)
+                end = torch.cuda.Event(enable_timing=True)
+                start.record()
+                for _ in range(10):
+                    custom_kernel(inputs)
+                end.record()
+                torch.cuda.synchronize()
+                impl_time_ms = start.elapsed_time(end) / 10
+
+                for _ in range(3):
+                    ref_kernel(inputs)
+                torch.cuda.synchronize()
+                start.record()
+                for _ in range(10):
+                    ref_kernel(inputs)
+                end.record()
+                torch.cuda.synchronize()
+                ref_time_ms = start.elapsed_time(end) / 10
 
             total_time_ms += impl_time_ms
             ref_total_time_ms += ref_time_ms
@@ -1503,6 +1812,23 @@ async def run_evaluate_workspace(
     ref_code = args.reference.read_text()
     test_cases_json = args.test_cases.read_text()
 
+    # Read defense module if defensive mode is enabled
+    defense_code = None
+    if args.defensive:
+        defense_path = (
+            Path(__file__).parent.parent.parent.parent
+            / "packages"
+            / "wafer-core"
+            / "wafer_core"
+            / "utils"
+            / "kernel_utils"
+            / "defense.py"
+        )
+        if defense_path.exists():
+            defense_code = defense_path.read_text()
+        else:
+            print(f"Warning: defense.py not found at {defense_path}, falling back to basic defense")
+
     # Build inline eval script
     eval_script = _build_workspace_eval_script(
         impl_code=impl_code,
@@ -1510,6 +1836,7 @@ async def run_evaluate_workspace(
         test_cases_json=test_cases_json,
         run_benchmarks=args.benchmark,
         run_defensive=args.defensive,
+        defense_code=defense_code,
     )
 
     # Execute via workspace exec
@@ -1855,15 +2182,12 @@ async def run_evaluate_runpod(
                 # Add venv bin to PATH so ninja (from pip) is found by torch.utils.cpp_extension
                 venv_bin = env_state.venv_bin
                 env_vars = f"PATH={venv_bin}:$PATH HIP_VISIBLE_DEVICES={gpu_id} ROCM_PATH=/opt/rocm"
-                pythonpath = f"PYTHONPATH={wafer_core_workspace}"
-                evaluate_script = (
-                    f"{wafer_core_workspace}/wafer_core/utils/kernel_utils/evaluate.py"
-                )
 
                 # Run from run_path so reference_kernel.py is importable
+                # Use installed wafer-core module
                 eval_cmd = (
                     f"cd {run_path} && "
-                    f"{env_vars} {pythonpath} {python_exe} {evaluate_script} "
+                    f"{env_vars} {python_exe} -m wafer_core.utils.kernel_utils.evaluate "
                     f"--implementation {impl_path} "
                     f"--reference {ref_path} "
                     f"--test-cases {test_cases_path} "
@@ -2219,15 +2543,12 @@ async def run_evaluate_digitalocean(
                     env_vars = (
                         f"PATH={venv_bin}:$PATH HIP_VISIBLE_DEVICES={gpu_id} ROCM_PATH=/opt/rocm"
                     )
-                    pythonpath = f"PYTHONPATH={wafer_core_workspace}"
-                    evaluate_script = (
-                        f"{wafer_core_workspace}/wafer_core/utils/kernel_utils/evaluate.py"
-                    )
 
                     # Run from run_path so reference_kernel.py is importable
+                    # Use installed wafer-core module
                     eval_cmd = (
                         f"cd {run_path} && "
-                        f"{env_vars} {pythonpath} {python_exe} {evaluate_script} "
+                        f"{env_vars} {python_exe} -m wafer_core.utils.kernel_utils.evaluate "
                         f"--implementation {impl_path} "
                         f"--reference {ref_path} "
                         f"--test-cases {test_cases_path} "
@@ -2407,7 +2728,9 @@ async def run_evaluate(args: EvaluateArgs) -> EvaluateResult:
     print(f"Using target: {target_name}")
 
     # Dispatch to appropriate executor
-    if isinstance(target, BaremetalTarget | VMTarget):
+    if isinstance(target, LocalTarget):
+        return await run_evaluate_local(args, target)
+    elif isinstance(target, BaremetalTarget | VMTarget):
         return await run_evaluate_ssh(args, target)
     elif isinstance(target, ModalTarget):
         return await run_evaluate_modal(args, target)
@@ -2436,6 +2759,7 @@ async def run_evaluate(args: EvaluateArgs) -> EvaluateResult:
 # Inline evaluation script for KernelBench format
 # This runs inside the Docker container on the remote GPU
 KERNELBENCH_EVAL_SCRIPT = """
+import gc
 import json
 import os
 import sys
@@ -2444,6 +2768,68 @@ import torch
 import torch.nn as nn
 from pathlib import Path
 
+# Use a unique per-run PyTorch extension cache directory to ensure fresh compilation.
+# This prevents stale cached extensions from being loaded when the pod is reused.
+# Without this, if a kernel is modified but uses the same extension name,
+# PyTorch would load the old cached .so instead of recompiling.
+# We use a UUID-based directory instead of clearing the cache to avoid race conditions
+# with other processes that might be using the cache.
+import uuid
+unique_cache_dir = f"/tmp/torch_extensions_{uuid.uuid4().hex[:8]}"
+os.environ["TORCH_EXTENSIONS_DIR"] = unique_cache_dir
+print(f"[KernelBench] Using unique extension cache: {unique_cache_dir}")
+
+# Clear any stale GPU memory from previous runs at startup
+# NOTE: empty_cache only frees memory from THIS process's PyTorch allocator.
+# It won't free memory from dead/zombie processes - rocm-smi --showpids can show
+# PIDs that no longer exist but still hold GPU memory. Those require a GPU reset
+# (rocm-smi --gpureset) to fully clear. TODO: detect and warn about orphaned memory.
+if torch.cuda.is_available():
+    gc.collect()
+    torch.cuda.empty_cache()
+    torch.cuda.reset_peak_memory_stats()
+
+
+def _calculate_timing_stats(times: list[float]) -> dict:
+    '''Calculate median and IQR from timing samples.
+
+    Returns dict with median, iqr_low (25th percentile), iqr_high (75th percentile),
+    mean, min, max, and std.
+    '''
+    import statistics
+
+    if not times:
+        return {"median": 0, "iqr_low": 0, "iqr_high": 0, "mean": 0, "min": 0, "max": 0, "std": 0}
+
+    sorted_times = sorted(times)
+    n = len(sorted_times)
+
+    # Median
+    median = statistics.median(sorted_times)
+
+    # Quartiles (25th and 75th percentile)
+    # For small samples, use simple interpolation
+    q1_idx = (n - 1) * 0.25
+    q3_idx = (n - 1) * 0.75
+
+    q1_low = int(q1_idx)
+    q1_frac = q1_idx - q1_low
+    iqr_low = sorted_times[q1_low] * (1 - q1_frac) + sorted_times[min(q1_low + 1, n - 1)] * q1_frac
+
+    q3_low = int(q3_idx)
+    q3_frac = q3_idx - q3_low
+    iqr_high = sorted_times[q3_low] * (1 - q3_frac) + sorted_times[min(q3_low + 1, n - 1)] * q3_frac
+
+    return {
+        "median": median,
+        "iqr_low": iqr_low,
+        "iqr_high": iqr_high,
+        "mean": statistics.mean(sorted_times),
+        "min": min(sorted_times),
+        "max": max(sorted_times),
+        "std": statistics.stdev(sorted_times) if n > 1 else 0,
+    }
+
 
 def run_profiling(model, inputs, name, output_dir):
     '''Run torch.profiler and return summary stats.'''
@@ -2674,12 +3060,26 @@ def main():
     parser.add_argument("--inputs", help="Custom inputs file to override get_inputs()/get_init_inputs()")
     parser.add_argument("--benchmark", action="store_true")
     parser.add_argument("--profile", action="store_true")
+    parser.add_argument("--defensive", action="store_true", help="Run full defense checks against reward hacking")
+    parser.add_argument("--defense-module", help="Path to defense.py module")
     parser.add_argument("--seed", type=int, default=42, help="Random seed for reproducibility")
     parser.add_argument("--num-correct-trials", type=int, default=3)
     parser.add_argument("--num-perf-trials", type=int, default=10)
     parser.add_argument("--output", required=True)
     args = parser.parse_args()
 
+    # Load defense module if defensive mode is enabled
+    defense_module = None
+    if args.defensive and args.defense_module:
+        try:
+            import importlib.util
+            defense_spec = importlib.util.spec_from_file_location("defense", args.defense_module)
+            defense_module = importlib.util.module_from_spec(defense_spec)
+            defense_spec.loader.exec_module(defense_module)
+            print("[KernelBench] Defense module loaded")
+        except Exception as e:
+            print(f"[KernelBench] Warning: Could not load defense module: {e}")
+
     # Create output directory for profiles
     output_dir = Path(args.output).parent
     profile_dir = output_dir / "profiles"
@@ -2813,47 +3213,102 @@ def main():
             inputs = get_inputs()
             inputs = [x.cuda() if isinstance(x, torch.Tensor) else x for x in inputs]
 
-            # Warmup
-            for _ in range(5):
-                with torch.no_grad():
-                    _ = new_model(*inputs)
-            torch.cuda.synchronize()
-
-            # Benchmark new model
-            start = torch.cuda.Event(enable_timing=True)
-            end = torch.cuda.Event(enable_timing=True)
-
-            times = []
-            for _ in range(args.num_perf_trials):
-                start.record()
-                with torch.no_grad():
-                    _ = new_model(*inputs)
-                end.record()
-                torch.cuda.synchronize()
-                times.append(start.elapsed_time(end))
-
-            new_time = sum(times) / len(times)
-            results["runtime_ms"] = new_time
-
-            # Benchmark reference model
-            for _ in range(5):
-                with torch.no_grad():
-                    _ = ref_model(*inputs)
-            torch.cuda.synchronize()
+            if args.defensive and defense_module is not None:
+                # Use full defense suite
+                print("[KernelBench] Running defense checks on implementation...")
+                run_all_defenses = defense_module.run_all_defenses
+                time_with_defenses = defense_module.time_execution_with_defenses
 
-            times = []
-            for _ in range(args.num_perf_trials):
-                start.record()
-                with torch.no_grad():
-                    _ = ref_model(*inputs)
-                end.record()
+                # Run defense checks on implementation
+                all_passed, defense_results, _ = run_all_defenses(
+                    lambda *x: new_model(*x),
+                    *inputs,
+                )
+                results["defense_results"] = {
+                    name: {"passed": passed, "message": msg}
+                    for name, passed, msg in defense_results
+                }
+                if not all_passed:
+                    failed = [name for name, passed, _ in defense_results if not passed]
+                    results["error"] = f"Defense checks failed: {failed}"
+                    print(f"[KernelBench] Defense checks FAILED: {failed}")
+                    for name, passed, msg in defense_results:
+                        status = "PASS" if passed else "FAIL"
+                        print(f"   [{status}] {name}: {msg}")
+                else:
+                    print("[KernelBench] All defense checks passed")
+
+                    # Time with defensive timing
+                    impl_times, _ = time_with_defenses(
+                        lambda: new_model(*inputs),
+                        [],
+                        num_warmup=5,
+                        num_trials=args.num_perf_trials,
+                        verbose=False,
+                        run_defenses=False,  # Already ran above
+                    )
+                    # Calculate stats for new model
+                    new_stats = _calculate_timing_stats(impl_times)
+                    results["runtime_ms"] = new_stats["median"]
+                    results["runtime_stats"] = new_stats
+
+                    # Reference timing
+                    ref_times, _ = time_with_defenses(
+                        lambda: ref_model(*inputs),
+                        [],
+                        num_warmup=5,
+                        num_trials=args.num_perf_trials,
+                        verbose=False,
+                        run_defenses=False,
+                    )
+                    ref_stats = _calculate_timing_stats(ref_times)
+                    results["reference_runtime_ms"] = ref_stats["median"]
+                    results["reference_runtime_stats"] = ref_stats
+                    results["speedup"] = ref_stats["median"] / new_stats["median"] if new_stats["median"] > 0 else 0
+                    print(f"[KernelBench] New: {new_stats['median']:.3f}ms (IQR: {new_stats['iqr_low']:.3f}-{new_stats['iqr_high']:.3f}), Ref: {ref_stats['median']:.3f}ms (IQR: {ref_stats['iqr_low']:.3f}-{ref_stats['iqr_high']:.3f}), Speedup: {results['speedup']:.2f}x")
+            else:
+                # Standard timing without full defenses
+                # Warmup BOTH models before benchmarking either
+                # This ensures consistent GPU state and avoids MIOpen cache effects
+                # that cause variance when warming up models sequentially
+                for _ in range(5):
+                    with torch.no_grad():
+                        _ = new_model(*inputs)
+                        _ = ref_model(*inputs)
                 torch.cuda.synchronize()
-                times.append(start.elapsed_time(end))
 
-            ref_time = sum(times) / len(times)
-            results["reference_runtime_ms"] = ref_time
-            results["speedup"] = ref_time / new_time if new_time > 0 else 0
-            print(f"[KernelBench] New: {new_time:.3f}ms, Ref: {ref_time:.3f}ms, Speedup: {results['speedup']:.2f}x")
+                # Benchmark new model
+                start = torch.cuda.Event(enable_timing=True)
+                end = torch.cuda.Event(enable_timing=True)
+
+                new_times = []
+                for _ in range(args.num_perf_trials):
+                    start.record()
+                    with torch.no_grad():
+                        _ = new_model(*inputs)
+                    end.record()
+                    torch.cuda.synchronize()
+                    new_times.append(start.elapsed_time(end))
+
+                new_stats = _calculate_timing_stats(new_times)
+                results["runtime_ms"] = new_stats["median"]
+                results["runtime_stats"] = new_stats
+
+                # Benchmark reference model
+                ref_times = []
+                for _ in range(args.num_perf_trials):
+                    start.record()
+                    with torch.no_grad():
+                        _ = ref_model(*inputs)
+                    end.record()
+                    torch.cuda.synchronize()
+                    ref_times.append(start.elapsed_time(end))
+
+                ref_stats = _calculate_timing_stats(ref_times)
+                results["reference_runtime_ms"] = ref_stats["median"]
+                results["reference_runtime_stats"] = ref_stats
+                results["speedup"] = ref_stats["median"] / new_stats["median"] if new_stats["median"] > 0 else 0
+                print(f"[KernelBench] New: {new_stats['median']:.3f}ms (IQR: {new_stats['iqr_low']:.3f}-{new_stats['iqr_high']:.3f}), Ref: {ref_stats['median']:.3f}ms (IQR: {ref_stats['iqr_low']:.3f}-{ref_stats['iqr_high']:.3f}), Speedup: {results['speedup']:.2f}x")
 
         # Run profiling if requested and correctness passed
         if args.profile and all_correct:
@@ -2898,6 +3353,16 @@ def main():
         json.dump(results, f, indent=2)
     print(f"[KernelBench] Results written to {args.output}")
 
+    # Cleanup GPU memory
+    try:
+        del ref_model, new_model
+    except NameError:
+        pass
+    import gc
+    gc.collect()
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
+
 if __name__ == "__main__":
     main()
 """
@@ -2947,6 +3412,27 @@ def _validate_kernelbench_files(args: KernelBenchEvaluateArgs) -> str | None:
             "  KernelBench format requires: 'class Model', 'get_inputs()', 'get_init_inputs()'"
         )
 
+    # Static kernel validation if backend specified
+    if args.backend:
+        from wafer_core.utils.kernel_utils.static_checker import validate_kernel_static
+
+        code = args.implementation.read_text()
+        valid, errors, warnings = validate_kernel_static(code, backend=args.backend)
+
+        # Print warnings (don't fail)
+        for warning in warnings:
+            logger.warning(f"Static check warning: {warning}")
+
+        # Fail on errors
+        if not valid:
+            error_list = "\n  - ".join(errors)
+            return (
+                f"Static kernel validation failed for backend '{args.backend}':\n"
+                f"  - {error_list}\n\n"
+                f"The implementation must use {args.backend.upper()} kernel primitives.\n"
+                "See KernelBench documentation for valid kernel patterns."
+            )
+
     return None
 
 
@@ -3059,6 +3545,30 @@ async def run_evaluate_kernelbench_docker(
                 error_message=f"Failed to write eval script: {write_result.stderr}",
             )
 
+        # Write defense module if defensive mode is enabled
+        defense_module_path = None
+        if args.defensive:
+            defense_path = (
+                Path(__file__).parent.parent.parent.parent
+                / "packages"
+                / "wafer-core"
+                / "wafer_core"
+                / "utils"
+                / "kernel_utils"
+                / "defense.py"
+            )
+            if defense_path.exists():
+                defense_code = defense_path.read_text()
+                defense_module_path = f"{run_path}/defense.py"
+                write_result = await client.exec(
+                    f"cat > '{defense_module_path}' << 'DEFENSE_EOF'\n{defense_code}\nDEFENSE_EOF"
+                )
+                if write_result.exit_code != 0:
+                    print(f"Warning: Failed to write defense module: {write_result.stderr}")
+                    defense_module_path = None
+            else:
+                print(f"Warning: defense.py not found at {defense_path}")
+
         print("Running KernelBench evaluation in Docker container...")
 
         # Paths inside container
@@ -3068,6 +3578,7 @@ async def run_evaluate_kernelbench_docker(
         container_inputs_path = f"{container_run_path}/custom_inputs.py" if args.inputs else None
         container_eval_script = f"{container_run_path}/kernelbench_eval.py"
         container_output = f"{container_run_path}/results.json"
+        container_defense_path = f"{container_run_path}/defense.py" if defense_module_path else None
 
         # Build eval command
         python_cmd_parts = [
@@ -3083,6 +3594,9 @@ async def run_evaluate_kernelbench_docker(
             python_cmd_parts.append("--profile")
         if container_inputs_path:
             python_cmd_parts.append(f"--inputs {container_inputs_path}")
+        if args.defensive and container_defense_path:
+            python_cmd_parts.append("--defensive")
+            python_cmd_parts.append(f"--defense-module {container_defense_path}")
         python_cmd_parts.append(f"--seed {args.seed}")
 
         eval_cmd = " ".join(python_cmd_parts)
@@ -3106,7 +3620,7 @@ async def run_evaluate_kernelbench_docker(
         # Run and stream output
         log_lines = []
         async for line in client.exec_stream(docker_cmd):
-            print(line)
+            print(line, flush=True)
             log_lines.append(line)
 
         # Read results
@@ -3298,15 +3812,44 @@ async def run_evaluate_kernelbench_digitalocean(
                         error_message=f"Failed to write eval script: {write_result.stderr}",
                     )
 
+                # Write defense module if defensive mode is enabled
+                defense_module_path = None
+                if args.defensive:
+                    defense_path = (
+                        Path(__file__).parent.parent.parent.parent
+                        / "packages"
+                        / "wafer-core"
+                        / "wafer_core"
+                        / "utils"
+                        / "kernel_utils"
+                        / "defense.py"
+                    )
+                    if defense_path.exists():
+                        defense_code = defense_path.read_text()
+                        defense_module_path = f"{run_path}/defense.py"
+                        write_result = await client.exec(
+                            f"cat > '{defense_module_path}' << 'DEFENSE_EOF'\n{defense_code}\nDEFENSE_EOF"
+                        )
+                        if write_result.exit_code != 0:
+                            print(f"Warning: Failed to write defense module: {write_result.stderr}")
+                            defense_module_path = None
+                    else:
+                        print(f"Warning: defense.py not found at {defense_path}")
+
                 print("Running KernelBench evaluation in Docker container (AMD/ROCm)...")
 
                 # Paths inside container
                 container_run_path = f"{CONTAINER_WORKSPACE}/{run_dir}"
                 container_impl_path = f"{container_run_path}/implementation.py"
                 container_ref_path = f"{container_run_path}/reference.py"
-                container_inputs_path = f"{container_run_path}/custom_inputs.py" if args.inputs else None
+                container_inputs_path = (
+                    f"{container_run_path}/custom_inputs.py" if args.inputs else None
+                )
                 container_eval_script = f"{container_run_path}/kernelbench_eval.py"
                 container_output = f"{container_run_path}/results.json"
+                container_defense_path = (
+                    f"{container_run_path}/defense.py" if defense_module_path else None
+                )
 
                 # Build eval command
                 python_cmd_parts = [
@@ -3322,6 +3865,9 @@ async def run_evaluate_kernelbench_digitalocean(
                     python_cmd_parts.append("--profile")
                 if container_inputs_path:
                     python_cmd_parts.append(f"--inputs {container_inputs_path}")
+                if args.defensive and container_defense_path:
+                    python_cmd_parts.append("--defensive")
+                    python_cmd_parts.append(f"--defense-module {container_defense_path}")
                 python_cmd_parts.append(f"--seed {args.seed}")
 
                 eval_cmd = " ".join(python_cmd_parts)
@@ -3330,14 +3876,20 @@ async def run_evaluate_kernelbench_digitalocean(
                 full_cmd = f"cd {container_run_path} && {eval_cmd}"
 
                 # Build Docker command for AMD
+                # PYTORCH_ROCM_ARCH: compile only for target arch (5-7x faster compile)
+                rocm_arch = _get_rocm_arch(target.compute_capability)
+                env_dict = {
+                    "HIP_VISIBLE_DEVICES": str(gpu_id),
+                    "PYTHONUNBUFFERED": "1",
+                }
+                if rocm_arch:
+                    env_dict["PYTORCH_ROCM_ARCH"] = rocm_arch
+
                 docker_cmd = _build_docker_run_command_amd(
                     image=docker_image,
                     command=full_cmd,
                     working_dir=container_run_path,
-                    env={
-                        "HIP_VISIBLE_DEVICES": str(gpu_id),
-                        "PYTHONUNBUFFERED": "1",
-                    },
+                    env=env_dict,
                     volumes={workspace_path: CONTAINER_WORKSPACE},
                 )
 
@@ -3346,7 +3898,7 @@ async def run_evaluate_kernelbench_digitalocean(
                 # Run and stream output
                 log_lines = []
                 async for line in client.exec_stream(docker_cmd):
-                    print(line)
+                    print(line, flush=True)
                     log_lines.append(line)
 
                 # Read results
@@ -3407,55 +3959,528 @@ async def run_evaluate_kernelbench_digitalocean(
                 )
 
 
-async def run_evaluate_kernelbench(args: KernelBenchEvaluateArgs) -> EvaluateResult:
-    """Run KernelBench format evaluation on configured target.
-
-    Args:
-        args: KernelBench evaluate arguments
+async def run_evaluate_kernelbench_runpod(
+    args: KernelBenchEvaluateArgs,
+    target: RunPodTarget,
+) -> EvaluateResult:
+    """Run KernelBench format evaluation directly on RunPod AMD GPU.
 
-    Returns:
-        Evaluation result
+    Runs evaluation script directly on host (no Docker) since RunPod pods
+    already have PyTorch/ROCm installed.
     """
-    from .targets import get_default_target, load_target
+    from datetime import datetime
 
-    # Validate input files
-    err = _validate_kernelbench_files(args)
-    if err:
-        return EvaluateResult(
-            success=False,
-            all_correct=False,
-            correctness_score=0.0,
-            geomean_speedup=0.0,
-            passed_tests=0,
-            total_tests=0,
-            error_message=err,
-        )
+    from wafer_core.async_ssh import AsyncSSHClient
+    from wafer_core.targets.runpod import RunPodError, runpod_ssh_context
 
-    # Load target
-    target_name = args.target_name
-    if not target_name:
-        target_name = get_default_target()
-        if not target_name:
-            return EvaluateResult(
-                success=False,
-                all_correct=False,
-                correctness_score=0.0,
-                geomean_speedup=0.0,
-                passed_tests=0,
-                total_tests=0,
-                error_message=(
-                    "No target specified and no default set.\n"
-                    "Set up a target first:\n"
-                    "  wafer config targets init ssh --name my-gpu --host user@host:22\n"
-                    "  wafer config targets init runpod --gpu MI300X\n"
-                    "Then use: --target my-gpu (or set default: wafer config targets default my-gpu)"
-                ),
-            )
+    REMOTE_WORKSPACE_BASE = "/tmp/wafer_eval"
+
+    # Select GPU
+    gpu_id = args.gpu_id if args.gpu_id is not None else target.gpu_ids[0]
+
+    print(f"Provisioning RunPod ({target.gpu_type_id})...")
 
     try:
-        target = load_target(target_name)
-    except FileNotFoundError:
-        return EvaluateResult(
+        async with runpod_ssh_context(target) as ssh_info:
+            ssh_target = f"{ssh_info.user}@{ssh_info.host}:{ssh_info.port}"
+            print(f"Connected to RunPod: {ssh_target}")
+
+            async with AsyncSSHClient(ssh_target, target.ssh_key) as client:
+                # Create workspace
+                timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+                run_dir = f"kernelbench_eval_{timestamp}"
+                run_path = f"{REMOTE_WORKSPACE_BASE}/{run_dir}"
+
+                await client.exec(f"mkdir -p {run_path}")
+                print(f"Created run directory: {run_path}")
+
+                # Read and upload files
+                impl_code = args.implementation.read_text()
+                ref_code = args.reference.read_text()
+
+                # Write implementation
+                impl_path = f"{run_path}/implementation.py"
+                write_result = await client.exec(
+                    f"cat > '{impl_path}' << 'IMPL_EOF'\n{impl_code}\nIMPL_EOF"
+                )
+                if write_result.exit_code != 0:
+                    return EvaluateResult(
+                        success=False,
+                        all_correct=False,
+                        correctness_score=0.0,
+                        geomean_speedup=0.0,
+                        passed_tests=0,
+                        total_tests=0,
+                        error_message=f"Failed to write implementation: {write_result.stderr}",
+                    )
+
+                # Write reference
+                ref_path = f"{run_path}/reference.py"
+                write_result = await client.exec(
+                    f"cat > '{ref_path}' << 'REF_EOF'\n{ref_code}\nREF_EOF"
+                )
+                if write_result.exit_code != 0:
+                    return EvaluateResult(
+                        success=False,
+                        all_correct=False,
+                        correctness_score=0.0,
+                        geomean_speedup=0.0,
+                        passed_tests=0,
+                        total_tests=0,
+                        error_message=f"Failed to write reference: {write_result.stderr}",
+                    )
+
+                # Write custom inputs if provided
+                inputs_path = None
+                if args.inputs:
+                    inputs_code = args.inputs.read_text()
+                    inputs_path = f"{run_path}/custom_inputs.py"
+                    write_result = await client.exec(
+                        f"cat > '{inputs_path}' << 'INPUTS_EOF'\n{inputs_code}\nINPUTS_EOF"
+                    )
+                    if write_result.exit_code != 0:
+                        return EvaluateResult(
+                            success=False,
+                            all_correct=False,
+                            correctness_score=0.0,
+                            geomean_speedup=0.0,
+                            passed_tests=0,
+                            total_tests=0,
+                            error_message=f"Failed to write custom inputs: {write_result.stderr}",
+                        )
+
+                # Write eval script
+                eval_script_path = f"{run_path}/kernelbench_eval.py"
+                write_result = await client.exec(
+                    f"cat > '{eval_script_path}' << 'EVAL_EOF'\n{KERNELBENCH_EVAL_SCRIPT}\nEVAL_EOF"
+                )
+                if write_result.exit_code != 0:
+                    return EvaluateResult(
+                        success=False,
+                        all_correct=False,
+                        correctness_score=0.0,
+                        geomean_speedup=0.0,
+                        passed_tests=0,
+                        total_tests=0,
+                        error_message=f"Failed to write eval script: {write_result.stderr}",
+                    )
+
+                # Write defense module if defensive mode is enabled
+                defense_module_path = None
+                if args.defensive:
+                    defense_path = (
+                        Path(__file__).parent.parent.parent.parent
+                        / "packages"
+                        / "wafer-core"
+                        / "wafer_core"
+                        / "utils"
+                        / "kernel_utils"
+                        / "defense.py"
+                    )
+                    if defense_path.exists():
+                        defense_code = defense_path.read_text()
+                        defense_module_path = f"{run_path}/defense.py"
+                        write_result = await client.exec(
+                            f"cat > '{defense_module_path}' << 'DEFENSE_EOF'\n{defense_code}\nDEFENSE_EOF"
+                        )
+                        if write_result.exit_code != 0:
+                            print(f"Warning: Failed to write defense module: {write_result.stderr}")
+                            defense_module_path = None
+                    else:
+                        print(f"Warning: defense.py not found at {defense_path}")
+
+                print("Running KernelBench evaluation (AMD/ROCm)...")
+
+                # Find Python with PyTorch - check common locations on RunPod
+                python_exe = "python3"
+                for candidate in [
+                    "/opt/conda/envs/py_3.10/bin/python3",
+                    "/opt/conda/bin/python3",
+                ]:
+                    check = await client.exec(
+                        f"{candidate} -c 'import torch' 2>/dev/null && echo OK"
+                    )
+                    if "OK" in check.stdout:
+                        python_exe = candidate
+                        print(f"Using Python: {python_exe}")
+                        break
+
+                # Build eval command - run directly on host
+                output_path = f"{run_path}/results.json"
+                python_cmd_parts = [
+                    f"{python_exe} {eval_script_path}",
+                    f"--impl {impl_path}",
+                    f"--reference {ref_path}",
+                    f"--output {output_path}",
+                ]
+
+                if args.benchmark:
+                    python_cmd_parts.append("--benchmark")
+                if args.profile:
+                    python_cmd_parts.append("--profile")
+                if inputs_path:
+                    python_cmd_parts.append(f"--inputs {inputs_path}")
+                if args.defensive and defense_module_path:
+                    python_cmd_parts.append("--defensive")
+                    python_cmd_parts.append(f"--defense-module {defense_module_path}")
+                python_cmd_parts.append(f"--seed {args.seed}")
+
+                eval_cmd = " ".join(python_cmd_parts)
+
+                # Set environment for AMD GPU and run
+                # PYTORCH_ROCM_ARCH: compile only for target arch (5-7x faster compile)
+                rocm_arch = _get_rocm_arch(target.compute_capability)
+                arch_env = f"PYTORCH_ROCM_ARCH={rocm_arch}" if rocm_arch else ""
+                env_vars = f"HIP_VISIBLE_DEVICES={gpu_id} ROCM_PATH=/opt/rocm PYTHONUNBUFFERED=1 {arch_env}"
+                full_cmd = f"cd {run_path} && {env_vars} {eval_cmd}"
+
+                # Run and stream output
+                log_lines = []
+                async for line in client.exec_stream(full_cmd):
+                    print(line, flush=True)
+                    log_lines.append(line)
+
+                # Read results
+                cat_result = await client.exec(f"cat {output_path}")
+
+                if cat_result.exit_code != 0:
+                    log_tail = "\n".join(log_lines[-50:])
+                    return EvaluateResult(
+                        success=False,
+                        all_correct=False,
+                        correctness_score=0.0,
+                        geomean_speedup=0.0,
+                        passed_tests=0,
+                        total_tests=0,
+                        error_message=f"Evaluation failed. Log tail:\n{log_tail}",
+                    )
+
+                # Parse results
+                try:
+                    results_data = json.loads(cat_result.stdout)
+                except json.JSONDecodeError as e:
+                    return EvaluateResult(
+                        success=False,
+                        all_correct=False,
+                        correctness_score=0.0,
+                        geomean_speedup=0.0,
+                        passed_tests=0,
+                        total_tests=0,
+                        error_message=f"Failed to parse results: {e}",
+                    )
+
+                # Convert to EvaluateResult
+                correct = results_data.get("correct", False)
+                speedup = results_data.get("speedup", 0.0) or 0.0
+                error = results_data.get("error")
+
+                if error:
+                    return EvaluateResult(
+                        success=False,
+                        all_correct=False,
+                        correctness_score=0.0,
+                        geomean_speedup=0.0,
+                        passed_tests=0,
+                        total_tests=1,
+                        error_message=error,
+                    )
+
+                return EvaluateResult(
+                    success=True,
+                    all_correct=correct,
+                    correctness_score=1.0 if correct else 0.0,
+                    geomean_speedup=speedup,
+                    passed_tests=1 if correct else 0,
+                    total_tests=1,
+                )
+
+    except RunPodError as e:
+        return EvaluateResult(
+            success=False,
+            all_correct=False,
+            correctness_score=0.0,
+            geomean_speedup=0.0,
+            passed_tests=0,
+            total_tests=0,
+            error_message=f"RunPod error: {e}",
+        )
+
+
+async def run_evaluate_kernelbench_baremetal_amd(
+    args: KernelBenchEvaluateArgs,
+    target: BaremetalTarget,
+) -> EvaluateResult:
+    """Run KernelBench format evaluation directly on AMD baremetal target.
+
+    Runs evaluation script directly on host (no Docker) for AMD GPUs
+    that have PyTorch/ROCm installed.
+    """
+    from datetime import datetime
+
+    from wafer_core.async_ssh import AsyncSSHClient
+
+    REMOTE_WORKSPACE_BASE = "/tmp/wafer_eval"
+
+    # Select GPU
+    gpu_id = args.gpu_id if args.gpu_id is not None else target.gpu_ids[0]
+
+    print(f"Connecting to {target.ssh_target}...")
+
+    async with AsyncSSHClient(target.ssh_target, target.ssh_key) as client:
+        # Create workspace
+        timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+        run_dir = f"kernelbench_eval_{timestamp}"
+        run_path = f"{REMOTE_WORKSPACE_BASE}/{run_dir}"
+
+        await client.exec(f"mkdir -p {run_path}")
+        print(f"Created run directory: {run_path}")
+
+        # Read and upload files
+        impl_code = args.implementation.read_text()
+        ref_code = args.reference.read_text()
+
+        # Write implementation
+        impl_path = f"{run_path}/implementation.py"
+        write_result = await client.exec(
+            f"cat > '{impl_path}' << 'IMPL_EOF'\n{impl_code}\nIMPL_EOF"
+        )
+        if write_result.exit_code != 0:
+            return EvaluateResult(
+                success=False,
+                all_correct=False,
+                correctness_score=0.0,
+                geomean_speedup=0.0,
+                passed_tests=0,
+                total_tests=0,
+                error_message=f"Failed to write implementation: {write_result.stderr}",
+            )
+
+        # Write reference
+        ref_path = f"{run_path}/reference.py"
+        write_result = await client.exec(f"cat > '{ref_path}' << 'REF_EOF'\n{ref_code}\nREF_EOF")
+        if write_result.exit_code != 0:
+            return EvaluateResult(
+                success=False,
+                all_correct=False,
+                correctness_score=0.0,
+                geomean_speedup=0.0,
+                passed_tests=0,
+                total_tests=0,
+                error_message=f"Failed to write reference: {write_result.stderr}",
+            )
+
+        # Write custom inputs if provided
+        inputs_path = None
+        if args.inputs:
+            inputs_code = args.inputs.read_text()
+            inputs_path = f"{run_path}/custom_inputs.py"
+            write_result = await client.exec(
+                f"cat > '{inputs_path}' << 'INPUTS_EOF'\n{inputs_code}\nINPUTS_EOF"
+            )
+            if write_result.exit_code != 0:
+                return EvaluateResult(
+                    success=False,
+                    all_correct=False,
+                    correctness_score=0.0,
+                    geomean_speedup=0.0,
+                    passed_tests=0,
+                    total_tests=0,
+                    error_message=f"Failed to write custom inputs: {write_result.stderr}",
+                )
+
+        # Write eval script
+        eval_script_path = f"{run_path}/kernelbench_eval.py"
+        write_result = await client.exec(
+            f"cat > '{eval_script_path}' << 'EVAL_EOF'\n{KERNELBENCH_EVAL_SCRIPT}\nEVAL_EOF"
+        )
+        if write_result.exit_code != 0:
+            return EvaluateResult(
+                success=False,
+                all_correct=False,
+                correctness_score=0.0,
+                geomean_speedup=0.0,
+                passed_tests=0,
+                total_tests=0,
+                error_message=f"Failed to write eval script: {write_result.stderr}",
+            )
+
+        # Write defense module if defensive mode is enabled
+        defense_module_path = None
+        if args.defensive:
+            defense_path = (
+                Path(__file__).parent.parent.parent.parent
+                / "packages"
+                / "wafer-core"
+                / "wafer_core"
+                / "utils"
+                / "kernel_utils"
+                / "defense.py"
+            )
+            if defense_path.exists():
+                defense_code = defense_path.read_text()
+                defense_module_path = f"{run_path}/defense.py"
+                write_result = await client.exec(
+                    f"cat > '{defense_module_path}' << 'DEFENSE_EOF'\n{defense_code}\nDEFENSE_EOF"
+                )
+                if write_result.exit_code != 0:
+                    print(f"Warning: Failed to write defense module: {write_result.stderr}")
+                    defense_module_path = None
+            else:
+                print(f"Warning: defense.py not found at {defense_path}")
+
+        print("Running KernelBench evaluation (AMD/ROCm)...")
+
+        # Find Python with PyTorch - check common locations
+        python_exe = "python3"
+        for candidate in [
+            "/opt/conda/envs/py_3.10/bin/python3",
+            "/opt/conda/bin/python3",
+        ]:
+            check = await client.exec(f"{candidate} -c 'import torch' 2>/dev/null && echo OK")
+            if "OK" in check.stdout:
+                python_exe = candidate
+                print(f"Using Python: {python_exe}")
+                break
+
+        # Build eval command - run directly on host
+        output_path = f"{run_path}/results.json"
+        python_cmd_parts = [
+            f"{python_exe} {eval_script_path}",
+            f"--impl {impl_path}",
+            f"--reference {ref_path}",
+            f"--output {output_path}",
+        ]
+
+        if args.benchmark:
+            python_cmd_parts.append("--benchmark")
+        if args.profile:
+            python_cmd_parts.append("--profile")
+        if inputs_path:
+            python_cmd_parts.append(f"--inputs {inputs_path}")
+        if args.defensive and defense_module_path:
+            python_cmd_parts.append("--defensive")
+            python_cmd_parts.append(f"--defense-module {defense_module_path}")
+        python_cmd_parts.append(f"--seed {args.seed}")
+
+        eval_cmd = " ".join(python_cmd_parts)
+
+        # Set environment for AMD GPU and run
+        # PYTORCH_ROCM_ARCH: compile only for target arch (5-7x faster compile)
+        rocm_arch = _get_rocm_arch(target.compute_capability)
+        arch_env = f"PYTORCH_ROCM_ARCH={rocm_arch}" if rocm_arch else ""
+        env_vars = f"HIP_VISIBLE_DEVICES={gpu_id} ROCM_PATH=/opt/rocm PYTHONUNBUFFERED=1 {arch_env}"
+        full_cmd = f"cd {run_path} && {env_vars} {eval_cmd}"
+
+        # Run and stream output
+        log_lines = []
+        async for line in client.exec_stream(full_cmd):
+            print(line, flush=True)
+            log_lines.append(line)
+
+        # Read results
+        cat_result = await client.exec(f"cat {output_path}")
+
+        if cat_result.exit_code != 0:
+            log_tail = "\n".join(log_lines[-50:])
+            return EvaluateResult(
+                success=False,
+                all_correct=False,
+                correctness_score=0.0,
+                geomean_speedup=0.0,
+                passed_tests=0,
+                total_tests=0,
+                error_message=f"Evaluation failed. Log tail:\n{log_tail}",
+            )
+
+        # Parse results
+        try:
+            results_data = json.loads(cat_result.stdout)
+        except json.JSONDecodeError as e:
+            return EvaluateResult(
+                success=False,
+                all_correct=False,
+                correctness_score=0.0,
+                geomean_speedup=0.0,
+                passed_tests=0,
+                total_tests=0,
+                error_message=f"Failed to parse results: {e}",
+            )
+
+        # Convert to EvaluateResult
+        correct = results_data.get("correct", False)
+        speedup = results_data.get("speedup", 0.0) or 0.0
+        error = results_data.get("error")
+
+        if error:
+            return EvaluateResult(
+                success=False,
+                all_correct=False,
+                correctness_score=0.0,
+                geomean_speedup=0.0,
+                passed_tests=0,
+                total_tests=1,
+                error_message=error,
+            )
+
+        return EvaluateResult(
+            success=True,
+            all_correct=correct,
+            correctness_score=1.0 if correct else 0.0,
+            geomean_speedup=speedup,
+            passed_tests=1 if correct else 0,
+            total_tests=1,
+        )
+
+
+async def run_evaluate_kernelbench(args: KernelBenchEvaluateArgs) -> EvaluateResult:
+    """Run KernelBench format evaluation on configured target.
+
+    Args:
+        args: KernelBench evaluate arguments
+
+    Returns:
+        Evaluation result
+    """
+    from .targets import get_default_target, load_target
+
+    # Validate input files
+    err = _validate_kernelbench_files(args)
+    if err:
+        return EvaluateResult(
+            success=False,
+            all_correct=False,
+            correctness_score=0.0,
+            geomean_speedup=0.0,
+            passed_tests=0,
+            total_tests=0,
+            error_message=err,
+        )
+
+    # Load target
+    target_name = args.target_name
+    if not target_name:
+        target_name = get_default_target()
+        if not target_name:
+            return EvaluateResult(
+                success=False,
+                all_correct=False,
+                correctness_score=0.0,
+                geomean_speedup=0.0,
+                passed_tests=0,
+                total_tests=0,
+                error_message=(
+                    "No target specified and no default set.\n"
+                    "Set up a target first:\n"
+                    "  wafer config targets init ssh --name my-gpu --host user@host:22\n"
+                    "  wafer config targets init runpod --gpu MI300X\n"
+                    "Then use: --target my-gpu (or set default: wafer config targets default my-gpu)"
+                ),
+            )
+
+    try:
+        target = load_target(target_name)
+    except FileNotFoundError:
+        return EvaluateResult(
             success=False,
             all_correct=False,
             correctness_score=0.0,
@@ -3471,7 +4496,13 @@ async def run_evaluate_kernelbench(args: KernelBenchEvaluateArgs) -> EvaluateRes
     if isinstance(target, DigitalOceanTarget):
         # DigitalOcean AMD MI300X - uses ROCm Docker with device passthrough
         return await run_evaluate_kernelbench_digitalocean(args, target)
+    elif isinstance(target, RunPodTarget):
+        # RunPod AMD MI300X - uses ROCm Docker with device passthrough
+        return await run_evaluate_kernelbench_runpod(args, target)
     elif isinstance(target, BaremetalTarget | VMTarget):
+        # Check if this is an AMD target (gfx* compute capability) - run directly
+        if target.compute_capability and target.compute_capability.startswith("gfx"):
+            return await run_evaluate_kernelbench_baremetal_amd(args, target)
         # NVIDIA targets - require docker_image to be set
         if not target.docker_image:
             return EvaluateResult(
@@ -3497,6 +4528,6 @@ async def run_evaluate_kernelbench(args: KernelBenchEvaluateArgs) -> EvaluateRes
             total_tests=0,
             error_message=(
                 f"Target type '{type(target).__name__}' not yet supported for KernelBench format. "
-                "Use a DigitalOcean, Baremetal, or VM target."
+                "Use a DigitalOcean, RunPod, Baremetal, or VM target."
             ),
         )