wafer-core 0.1.21__py3-none-any.whl → 0.1.23__py3-none-any.whl

wafer_core/utils/kernel_utils/static_checker.py

@@ -155,6 +155,161 @@ def check_torch_computation_ops(code: str) -> tuple[bool, str]:
     return (False, "")
 
 
+# =============================================================================
+# NN.MODULE FORWARD CALL CHECKS (Reward Hacking in forward())
+# =============================================================================
+
+# These patterns detect calling PyTorch nn.Module forward methods inside forward()
+# e.g., self.conv(x), self.linear(x), self.bn(x) - these invoke cuBLAS/cuDNN
+# 
+# This is different from:
+# - nn.Linear(...) in __init__ = OK (just creates parameter container)
+# - self.linear.weight in forward() = OK (accessing weights for custom kernel)
+# - self.linear(x) in forward() = BAD (invokes PyTorch's matmul via cuBLAS)
+
+NN_MODULE_FORWARD_PATTERNS = [
+    # Common layer types being called as functions
+    r"self\.(conv\d*d?|linear|bn|batch_norm|layer_norm|group_norm|instance_norm)\s*\(",
+    # More generic pattern: self.<name>(x) or self.<name>(input)
+    # But we need to be careful not to match custom module calls
+]
+
+# =============================================================================
+# TORCH.NN.FUNCTIONAL CHECKS (Reward Hacking)
+# =============================================================================
+
+# Patterns for torch.nn.functional / F.* calls that bypass custom kernel requirement
+# These call into cuBLAS/cuDNN under the hood
+TORCH_FUNCTIONAL_PATTERNS = [
+    # F.linear, F.conv*, F.batch_norm etc. (common alias)
+    r"\bF\.(linear|conv[123]d|conv_transpose[123]d|batch_norm|layer_norm|group_norm|instance_norm)\s*\(",
+    # Full path torch.nn.functional.*
+    r"\btorch\.nn\.functional\.(linear|conv[123]d|conv_transpose[123]d|batch_norm|layer_norm|group_norm|instance_norm)\s*\(",
+]
+
+
+def check_torch_functional_calls(code: str) -> tuple[bool, str]:
+    """Check for torch.nn.functional / F.* calls in forward() method (reward hacking).
+    
+    Detects patterns like F.linear(x, weight), F.conv2d(x, weight) which invoke
+    PyTorch's built-in operations (backed by cuBLAS/cuDNN) instead of custom kernels.
+    
+    Note: We only check inside forward() to avoid false positives from imports or
+    training-mode fallbacks that aren't used during inference.
+    """
+    # Only check inside forward() method
+    forward_code = _extract_forward_method(code)
+    if not forward_code:
+        return (False, "")
+    
+    forward_code = _strip_comments(forward_code)
+    
+    for pattern in TORCH_FUNCTIONAL_PATTERNS:
+        match = re.search(pattern, forward_code)
+        if match:
+            return (True, f"Uses torch.nn.functional in forward(): {match.group(0)} (reward hacking - must use custom kernel)")
+    
+    return (False, "")
+
+
+def _extract_forward_method(code: str) -> str:
+    """Extract only the forward() method content from code."""
+    lines = code.split('\n')
+    result = []
+    in_forward = False
+    base_indent = 0
+    
+    for i, line in enumerate(lines):
+        # Look for forward method definition
+        if re.search(r'\bdef\s+forward\s*\(\s*self', line):
+            in_forward = True
+            # Get the indentation level of the def line
+            base_indent = len(line) - len(line.lstrip())
+            result.append(line)
+            continue
+        
+        if in_forward:
+            # Check if we've exited the forward method (new method/class at same or lower indent)
+            stripped = line.strip()
+            if stripped and not stripped.startswith('#') and not stripped.startswith('"""') and not stripped.startswith("'''"):
+                current_indent = len(line) - len(line.lstrip())
+                # If we hit a new def/class at the same or lower indentation, we're done
+                if current_indent <= base_indent and (stripped.startswith('def ') or stripped.startswith('class ')):
+                    break
+            result.append(line)
+    
+    return '\n'.join(result)
+
+
+def check_nn_module_forward_call(code: str) -> tuple[bool, str]:
+    """Check for nn.Module forward calls inside forward() method (reward hacking).
+    
+    Detects patterns like self.conv(x), self.linear(x) which invoke PyTorch's
+    built-in layers (backed by cuBLAS/cuDNN) instead of custom kernels.
+    """
+    # Only check inside forward() method
+    forward_code = _extract_forward_method(code)
+    if not forward_code:
+        return (False, "")
+    
+    forward_code = _strip_comments(forward_code)
+    
+    for pattern in NN_MODULE_FORWARD_PATTERNS:
+        match = re.search(pattern, forward_code)
+        if match:
+            return (True, f"Calls PyTorch nn.Module in forward(): {match.group(0)} (reward hacking - must use custom kernel)")
+    
+    return (False, "")
+
+
+# =============================================================================
+# CUBLAS/CUDNN DIRECT USAGE CHECKS (Reward Hacking)
+# =============================================================================
+
+# Direct cuBLAS calls bypass custom kernel requirement
+CUBLAS_PATTERNS = [
+    r"\bcublas[A-Z]\w+\s*\(",  # cublasSgemm, cublasGemmEx, etc.
+    r"\bcublasCreate\b",
+    r"\bcublasDestroy\b",
+    r"\bcublasSetStream\b",
+    r"\bcublasSetMathMode\b",
+    r"#include\s*[<\"]cublas",  # #include <cublas_v2.h>
+    r"CUBLAS_TENSOR_OP_MATH",
+]
+
+# Direct cuDNN calls bypass custom kernel requirement
+CUDNN_PATTERNS = [
+    r"\bcudnn[A-Z]\w+\s*\(",  # cudnnConvolutionForward, etc.
+    r"\bcudnnCreate\b",
+    r"\bcudnnDestroy\b",
+    r"#include\s*[<\"]cudnn",  # #include <cudnn.h>
+]
+
+
+def check_cublas_usage(code: str) -> tuple[bool, str]:
+    """Check for direct cuBLAS API usage (reward hacking)."""
+    code = _strip_comments(code)
+    
+    for pattern in CUBLAS_PATTERNS:
+        match = re.search(pattern, code)
+        if match:
+            return (True, f"Uses cuBLAS directly: {match.group(0)} (reward hacking - must write custom kernel)")
+    
+    return (False, "")
+
+
+def check_cudnn_usage(code: str) -> tuple[bool, str]:
+    """Check for direct cuDNN API usage (reward hacking)."""
+    code = _strip_comments(code)
+    
+    for pattern in CUDNN_PATTERNS:
+        match = re.search(pattern, code)
+        if match:
+            return (True, f"Uses cuDNN directly: {match.group(0)} (reward hacking - must write custom kernel)")
+    
+    return (False, "")
+
+
 # =============================================================================
 # TIMING MANIPULATION CHECKS
 # =============================================================================
@@ -310,8 +465,16 @@ def check_tk_impl(code: str) -> tuple[bool, str]:
 def check_cute_impl(code: str) -> tuple[bool, str]:
     """Check for valid CUTLASS/CuTe kernel implementation."""
     code = _strip_comments(code)
-    if not any(p in code for p in ["cute::", "cutlass::", "from cutlass"]):
-        return (True, "Missing cute:: or cutlass:: namespace")
+    # Accept explicit namespace qualifiers OR using namespace declarations
+    valid_patterns = [
+        "cute::",
+        "cutlass::",
+        "from cutlass",
+        "using namespace cute",
+        "using namespace cutlass",
+    ]
+    if not any(p in code for p in valid_patterns):
+        return (True, "Missing cute:: or cutlass:: namespace (or 'using namespace')")
     return (False, "")
 
 
@@ -334,6 +497,11 @@ CHECK_FUNCTIONS: dict[str, Callable[[str], tuple[bool, str]]] = {
     "timing_event_patch": check_timing_event_patch,
     # Torch ops
     "torch_computation_ops": check_torch_computation_ops,
+    # Reward hacking checks
+    "cublas_usage": check_cublas_usage,
+    "cudnn_usage": check_cudnn_usage,
+    "nn_module_forward_call": check_nn_module_forward_call,
+    "torch_functional_calls": check_torch_functional_calls,
     # Timing manipulation
     "stream_injection": check_stream_injection,
     "thread_injection": check_thread_injection,
@@ -363,12 +531,16 @@ STRICT_CHECKS = [
     "timing_event_patch",
     "thread_injection",
     "lazy_eval",
+    "cublas_usage",  # Direct cuBLAS is reward hacking
+    "cudnn_usage",   # Direct cuDNN is reward hacking
+    "nn_module_forward_call",  # Calling self.conv(x), self.linear(x) in forward() is reward hacking
+    "torch_functional_calls",  # Calling F.linear(), F.conv2d() in forward() is reward hacking
+    "torch_computation_ops",  # torch.mm, torch.matmul, torch.conv* etc. are reward hacking
 ]
 
 # Checks that emit warnings but don't fail
 WARNING_CHECKS = [
     "pytorch_wrap",
-    "torch_computation_ops",
     "stream_injection",
 ]
 

wafer_core/utils/kernel_utils/targets/config.py

@@ -61,7 +61,9 @@ class BaremetalTarget:
     ncu_available: bool = True  # Baremetal typically has NCU
 
     # Docker execution config (Modal-like). If docker_image is set, run in container.
-    docker_image: str | None = None  # Docker image to use (e.g., "nvcr.io/nvidia/cutlass:4.3-devel")
+    docker_image: str | None = (
+        None  # Docker image to use (e.g., "nvcr.io/nvidia/cutlass:4.3-devel")
+    )
     pip_packages: tuple[str, ...] = ()  # Packages to install via uv pip install
     torch_package: str | None = None  # Torch package spec (e.g., "torch>=2.8.0")
     torch_index_url: str | None = None  # Custom index for torch (e.g., PyTorch nightly)
@@ -69,7 +71,9 @@ class BaremetalTarget:
     def __post_init__(self) -> None:
         """Validate configuration."""
         assert len(self.gpu_ids) > 0, "Must specify at least one GPU ID"
-        assert ":" in self.ssh_target, f"ssh_target must include port (user@host:port), got: {self.ssh_target}"
+        assert ":" in self.ssh_target, (
+            f"ssh_target must include port (user@host:port), got: {self.ssh_target}"
+        )
         # If torch_index_url is set, torch_package must also be set
         if self.torch_index_url:
             assert self.torch_package, "torch_package must be set when torch_index_url is provided"
@@ -114,7 +118,9 @@ class VMTarget:
     ncu_available: bool = False  # VMs typically don't have NCU
 
     # Docker execution config (Modal-like). If docker_image is set, run in container.
-    docker_image: str | None = None  # Docker image to use (e.g., "nvcr.io/nvidia/pytorch:24.01-py3")
+    docker_image: str | None = (
+        None  # Docker image to use (e.g., "nvcr.io/nvidia/pytorch:24.01-py3")
+    )
     pip_packages: tuple[str, ...] = ()  # Packages to install via uv pip install
     torch_package: str | None = None  # Torch package spec (e.g., "torch>=2.8.0")
     torch_index_url: str | None = None  # Custom index for torch (e.g., PyTorch nightly)
@@ -122,7 +128,9 @@ class VMTarget:
     def __post_init__(self) -> None:
         """Validate configuration."""
         assert len(self.gpu_ids) > 0, "Must specify at least one GPU ID"
-        assert ":" in self.ssh_target, f"ssh_target must include port (user@host:port), got: {self.ssh_target}"
+        assert ":" in self.ssh_target, (
+            f"ssh_target must include port (user@host:port), got: {self.ssh_target}"
+        )
         # If torch_index_url is set, torch_package must also be set
         if self.torch_index_url:
             assert self.torch_package, "torch_package must be set when torch_index_url is provided"
@@ -282,7 +290,7 @@ class RunPodTarget:
             ssh_key="~/.ssh/id_ed25519",
             gpu_type_id="AMD Instinct MI300X OAM",
             gpu_count=1,
-            image="runpod/pytorch:2.4.0-py3.10-rocm6.1.0-ubuntu22.04",
+            image="rocm/pytorch:rocm7.0.2_ubuntu24.04_py3.12_pytorch_release_2.7.1",
             keep_alive=True,  # Don't terminate after eval
         )
 
@@ -296,7 +304,21 @@ class RunPodTarget:
     gpu_type_id: str = AMD_MI300X_GPU_ID  # RunPod GPU type identifier
     gpu_count: int = 1
     container_disk_gb: int = 50
-    image: str = "runpod/pytorch:2.4.0-py3.10-rocm6.1.0-ubuntu22.04"
+    # TODO: Consider creating a custom Docker image with HipKittens pre-installed
+    # to avoid needing `wafer config targets install <target> hipkittens`.
+    # HipKittens repo: https://github.com/HazyResearch/hipkittens
+    # CK (Composable Kernel) is already included in ROCm 7.0.
+    #
+    # WARNING: PyTorch's hipify can corrupt /opt/rocm/include/thrust/ headers.
+    # If you see "cuda/__cccl_config not found" errors, run:
+    #   apt-get install --reinstall -y rocthrust
+    # See docker/rocm7-runpod/README.md for details.
+    image: str = "rocm/pytorch:rocm7.0.2_ubuntu24.04_py3.12_pytorch_release_2.7.1"
+
+    # RunPod template ID — required for non-RunPod images that need custom
+    # dockerArgs (e.g. to install and start sshd). When set, takes priority
+    # over `image` in the deploy mutation.
+    template_id: str | None = None
 
     # Timeouts
     provision_timeout: int = 900  # 15 min for SSH to be ready
@@ -325,12 +347,13 @@ class RunPodTarget:
 
         # Check for API key (env var or ~/.wafer/auth.json)
         api_key = get_api_key("runpod")
-        assert api_key, (
-            "RunPod API key not found.\n"
-            "Set WAFER_RUNPOD_API_KEY environment variable, or run:\n"
-            "  wafer auth login runpod\n"
-            "Get your API key from: https://runpod.io/console/user/settings"
-        )
+        if not api_key:
+            raise ValueError(
+                "RunPod API key not found.\n"
+                "Set WAFER_RUNPOD_API_KEY environment variable, or run:\n"
+                "  wafer auth login runpod\n"
+                "Get your API key from: https://runpod.io/console/user/settings"
+            )
 
 
 @dataclass(frozen=True)
@@ -370,7 +393,9 @@ class LocalTarget:
         """Validate configuration."""
         assert self.name, "name cannot be empty"
         assert len(self.gpu_ids) > 0, "Must specify at least one GPU ID"
-        assert self.vendor in ("nvidia", "amd"), f"vendor must be 'nvidia' or 'amd', got: {self.vendor}"
+        assert self.vendor in ("nvidia", "amd"), (
+            f"vendor must be 'nvidia' or 'amd', got: {self.vendor}"
+        )
 
 
 @dataclass(frozen=True)
@@ -415,7 +440,7 @@ class DigitalOceanTarget:
     # DigitalOcean instance configuration
     region: str = "atl1"  # Atlanta (AMD GPUs available here)
     size_slug: str = "gpu-mi300x1-192gb-devcloud"  # Single MI300X GPU
-    image: str = "gpu-amd-base"  # AMD AI/ML Ready image with ROCm
+    image: str = "amd-pytorchrocm7"  # PyTorch (ROCm7) marketplace image
 
     # Timeouts
     provision_timeout: int = 600  # 10 min for droplet to be ready
@@ -443,16 +468,25 @@ class DigitalOceanTarget:
 
         # Check for API key (env var or ~/.wafer/auth.json)
         api_key = get_api_key("digitalocean")
-        assert api_key, (
-            "DigitalOcean API key not found.\n"
-            "Set WAFER_AMD_DIGITALOCEAN_API_KEY environment variable, or run:\n"
-            "  wafer auth login digitalocean\n"
-            "Get your API key from: https://cloud.digitalocean.com/account/api/tokens"
-        )
+        if not api_key:
+            raise ValueError(
+                "DigitalOcean API key not found.\n"
+                "Set WAFER_AMD_DIGITALOCEAN_API_KEY environment variable, or run:\n"
+                "  wafer auth login digitalocean\n"
+                "Get your API key from: https://cloud.digitalocean.com/account/api/tokens"
+            )
 
 
 # Union type for target configs
-TargetConfig = BaremetalTarget | VMTarget | ModalTarget | WorkspaceTarget | RunPodTarget | DigitalOceanTarget | LocalTarget
+TargetConfig = (
+    BaremetalTarget
+    | VMTarget
+    | ModalTarget
+    | WorkspaceTarget
+    | RunPodTarget
+    | DigitalOceanTarget
+    | LocalTarget
+)
 
 
 # Type guard functions for pattern matching
@@ -517,9 +551,9 @@ def target_to_deployment_config(target: TargetConfig, gpu_id: int) -> Deployment
     from wafer_core.utils.kernel_utils.deployment import DeploymentConfig
 
     # Type narrowing: Only SSH-based targets supported (not Modal)
-    assert not isinstance(
-        target, ModalTarget
-    ), f"target_to_deployment_config only supports SSH targets, got {type(target).__name__}"
+    assert not isinstance(target, ModalTarget), (
+        f"target_to_deployment_config only supports SSH targets, got {type(target).__name__}"
+    )
 
     return DeploymentConfig(
         ssh_target=target.ssh_target,

wafer_core/utils/modal_execution/modal_app.py

@@ -20,35 +20,17 @@ import modal
 
 # Build Modal image with all dependencies
 # This image is cached and reused across function invocations
-def build_modal_image(
-    gpu_type: str = "B200",
-    compute_capability: str = "10.0",
-) -> modal.Image:
+def build_modal_image() -> modal.Image:
     """Build Modal image with PyTorch, CUTLASS, and evaluation dependencies.
 
     Uses explicit local code inclusion to avoid pulling in SSH deployment code.
 
-    Phase 2 solution from MODAL_HANDOFF.md:
-    - Use add_local_dir with ignore parameter to exclude deployment files
-    - Only include files needed for kernel evaluation
-
-    Args:
-        gpu_type: GPU type (determines PyTorch index URL)
-        compute_capability: CUDA compute capability
-
     Returns:
         Modal Image ready for kernel evaluation
     """
-    # Determine PyTorch index based on GPU type
-    # Match logic from configs/base_config.py
-    if gpu_type in ["B200", "GB200"] or compute_capability.startswith("10."):
-        # Blackwell requires PyTorch 2.8+ with CUDA 12.8
-        torch_index = "https://download.pytorch.org/whl/nightly/cu128"
-        torch_version = "torch>=2.8.0"
-    else:
-        # Older GPUs (H100, A100) use stable PyTorch
-        torch_index = "https://download.pytorch.org/whl/cu124"
-        torch_version = "torch>=2.4.0"
+    # Use CUDA 13.0 for all GPUs (H100, A100, B200, GB200)
+    torch_index = "https://download.pytorch.org/whl/cu130"
+    torch_version = "torch>=2.6.0"
 
     # Build image with dependencies
     image = (
@@ -74,6 +56,15 @@ def build_modal_image(
             "scipy",
             "pytest",
         )
+        # Install CUTLASS headers for C++ kernel compilation (v4.3.5)
+        .run_commands(
+            "git clone --depth 1 --branch v4.3.5 https://github.com/NVIDIA/cutlass.git /usr/local/cutlass",
+            # Verify CUTLASS was installed correctly
+            "ls -la /usr/local/cutlass/include/cutlass/util/ | head -20",
+            "test -f /usr/local/cutlass/include/cutlass/util/packed_stride.hpp && echo 'CUTLASS headers OK' || echo 'CUTLASS headers MISSING'",
+        )
+        # Set CUTLASS_PATH environment variable
+        .env({"CUTLASS_PATH": "/usr/local/cutlass/include"})
         # Create empty __init__.py files for proper Python package structure
         # MUST run before add_local_* commands (Modal restriction)
         .run_commands(
@@ -111,20 +102,16 @@ def build_modal_image(
 # Create app (can be customized per target)
 def create_modal_app(
     app_name: str = "test-kernel-eval",  # Match test script default
-    gpu_type: str = "B200",
-    compute_capability: str = "10.0",
 ) -> modal.App:
     """Create Modal app for kernel evaluation.
 
     Args:
         app_name: Modal app name
-        gpu_type: GPU type for image building
-        compute_capability: CUDA compute capability
 
     Returns:
         Modal App instance
     """
-    image = build_modal_image(gpu_type=gpu_type, compute_capability=compute_capability)
+    image = build_modal_image()
     return modal.App(name=app_name, image=image)
 
 

{wafer_core-0.1.21.dist-info → wafer_core-0.1.23.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: wafer-core
-Version: 0.1.21
+Version: 0.1.23
 Summary: Core utilities and environments for Wafer GPU kernel optimization
 Requires-Python: >=3.10
 Requires-Dist: aiohttp>=3.9.0

{wafer_core-0.1.21.dist-info → wafer_core-0.1.23.dist-info}/RECORD

@@ -1,6 +1,6 @@
 wafer_core/__init__.py,sha256=syB2JjzvL91otODCKsGCgCPE6gVCINhTUv9j98k6tL0,3209
 wafer_core/async_ssh.py,sha256=ocw2Gh5p8ltKeoqG_q32DXOBfu5q-IE7jCnzMbQN9WI,28713
-wafer_core/auth.py,sha256=U1AeQE_Z5ev0ltXAEIh0STFIv46mie9G8Vy34oOyowE,5936
+wafer_core/auth.py,sha256=JpUkZ3bROIsgexayak5TLiGqUAR5kqGjekwqQRvIXH0,7235
 wafer_core/gpu.py,sha256=ENa92btjXsx6ldpoyKfRrAmfy-LHG2KpA5k7SWd6Q_s,28627
 wafer_core/gpu_detect.py,sha256=kpD8Q_G6GA9j-WnnnTNA3BBPulkGcWnZiogOmjKDao0,13650
 wafer_core/problem_config.py,sha256=8oqxL9-pvgzi8BtFxgDcqZW4e6DV2OCZOYkcPoyXrc8,10913
@@ -12,7 +12,7 @@ wafer_core/config/__init__.py,sha256=hKywfjA4YXd4lBeBFEcBoMwFoflPHJTiBnkTq7_JYOQ
 wafer_core/config/loader.py,sha256=k7JnILmO13TWUzIv9Lm8fvmj3UfYHZDgaFurjQ-GXpY,6623
 wafer_core/config/schema.py,sha256=2WhFlnG0VYYX4T-70BLeJK8Janvi4KEa8KKGZA7331w,3898
 wafer_core/environments/__init__.py,sha256=SIsResVtm22tr_d-oHPeeSxrkhFdmPOFico3DqtRqK8,238
-wafer_core/environments/coding.py,sha256=0TU1kggKH4fLaoeDOUKpWwBklDf5Q9dHTAV2b96GVnU,8313
+wafer_core/environments/coding.py,sha256=N-ELZwJu5vKLCVtwO25c6JSty6fmqf85VR2d3WJ4RXw,8559
 wafer_core/environments/gpumode.py,sha256=8Da08nltvN_YloNyYI6-omN2D4n5C7aptKDCtUgT2bQ,17191
 wafer_core/lib/__init__.py,sha256=4-4p3mhwlquejWGglYXU8_nHdA0LoPaa_jGzcm13USA,1325
 wafer_core/lib/kernel_scope/__init__.py,sha256=WW2vu8jUlqOu-MCpgO40lIYacCA9N2u-uuECIs_JO2w,2817
@@ -336,9 +336,9 @@ wafer_core/rollouts/agents.py,sha256=Uv1kjYogUfdPl18YfkVxVqFTbmWfuJQrxem_iHTUgdw
 wafer_core/rollouts/cli.py,sha256=2NqgegKdlmxD0eJzGOMB5o_1Hb5t7O5JpP_32uvF2BE,80117
 wafer_core/rollouts/cli_agents.py,sha256=e4qqqYBzWLsbw8FsNnddGApWp_on9Cvzrfd1amiAyvI,20641
 wafer_core/rollouts/deploy.py,sha256=3t88fM_BMyAPkxIl8pS4r5ogHJvrlqWQDuIaltDZBRc,40924
-wafer_core/rollouts/dtypes.py,sha256=GXB3SaLzlSXs3Gxe6GLdU3_t4E1Pl7EmJamexRuaFDE,60962
+wafer_core/rollouts/dtypes.py,sha256=GUezPTzkd8E-nDlqdGE7idUthyZC-7jTrbpa4ye-v8k,61146
 wafer_core/rollouts/eval_helpers.py,sha256=OE7uQZRcbqQhpFqb4zOj8zafc9Gr6xZJpSrMvxXKVUw,1699
-wafer_core/rollouts/evaluation.py,sha256=Zu2hqkGWDjTRj9UxRVgNQ_FgEnFo6zKCw4ck2htvnpg,68483
+wafer_core/rollouts/evaluation.py,sha256=fk-pGZ5vpocVmw1iBbHtxMK0K6l8pYTLHCpDNvRY1Xo,69142
 wafer_core/rollouts/events.py,sha256=z85J8kq0LXPj5CiUk4RkiTQg--r9xiO7QeeJwkyUOto,7505
 wafer_core/rollouts/export.py,sha256=0CfdBB7Du4E3VekKEUcTwTEFS1bOMGZ9GbD5KU3CecQ,11583
 wafer_core/rollouts/feedback.py,sha256=mu17eQbAinXZWI3hMYLq-LyF4JAdH9SfNRWY-0S8jvQ,6769
@@ -350,11 +350,12 @@ wafer_core/rollouts/paths.py,sha256=9XtrA9ylhb5LttMFe2DE7X0IHeUMjuGUerII9OscYec,
 wafer_core/rollouts/pipeline.py,sha256=vlJTYE3ZX2XScpF9pmtv91K8Q0g8uLmcbI5jn6b5Hzg,15319
 wafer_core/rollouts/progress.py,sha256=szA9cvWT2xUxGVhF9BaAqJMmKDqMAUlxImxcOpcnqbY,29228
 wafer_core/rollouts/progress_display.py,sha256=it-IiI37k9whAuB6T_66GYgsZyidCq5x00URiOcxe2c,15769
-wafer_core/rollouts/prompt.py,sha256=KAFSS4HCtYtaPGGCFtC5VU7NrVmhEoQFNGWcA1HVW-Q,10566
+wafer_core/rollouts/prompt.py,sha256=EDmGb0rhWwke7tokIcO8dukc3q5c8x0n5Omi5CpAQmA,11022
 wafer_core/rollouts/providers.py,sha256=dcGJh1p30hstVbCDDtJ902lyafkg81DKjcOzb0uuKS0,1400
 wafer_core/rollouts/remote.py,sha256=cAYpRCONlsTeRxzLiegAUfjZWGtqBNwZTHehMhk5ldA,8816
 wafer_core/rollouts/scoring.py,sha256=qeIT8Z7pK51XRDmN2sGdg_hIPRabWqoQIYKsuytlvRo,8838
 wafer_core/rollouts/search.py,sha256=5BEDuw9FVbQhei3nvUXEVwBU5ouwgJE6ONhEqvU5Ldc,14696
+wafer_core/rollouts/skills.py,sha256=ATYoG02Cc6_VrtE415TnseBFJrKOMq27z-5YgBgPpZQ,5081
 wafer_core/rollouts/slice.py,sha256=darOZO53BuSPfvv_KjOSzulGVSWbL4OuoE3k6xXpBFg,20195
 wafer_core/rollouts/store.py,sha256=UDP9idDOEVs_0Pslx0K_Y8E1i-BeoqVSaxdQiaqtz1E,18051
 wafer_core/rollouts/transform_messages.py,sha256=yldzdLgugNYb5Zxju7myFBel1tmrHXx9M399ImqPLGI,20891
@@ -394,7 +395,7 @@ wafer_core/rollouts/environments/compose.py,sha256=DlJA_GdzByWjVvGeR4MrcQIB4ucV6
 wafer_core/rollouts/environments/cuda_grep.py,sha256=o4GPJcnKuB96KwE4UWkxXq5DdaKYMz-QizcZXleOKLs,22007
 wafer_core/rollouts/environments/git_worktree.py,sha256=f4S-OI-m6OEMyrEl3TfD4oYLXkkNgsiHuX6NOHCVfSQ,22397
 wafer_core/rollouts/environments/handoff.py,sha256=pvhcSDdltZ1zJ3Y_SAJlBmX6FhXjZLNYneXwjFqb9lE,16941
-wafer_core/rollouts/environments/localfs.py,sha256=fpk91XB9GNrtsHiFhSzvkfzRdEyqf8me_pR5PPxDmMg,43848
+wafer_core/rollouts/environments/localfs.py,sha256=xceepFc-eh9RqwTSJd_WiODVCs4Jm-d8rQm_MvutG0A,45766
 wafer_core/rollouts/environments/no_tools.py,sha256=FRsiMDzi8ma15xePTANteUxAOiOw0s90459XMvSq1_k,2970
 wafer_core/rollouts/environments/oracle.py,sha256=OtviDKed6DEIy67TdohqSGmvQT8XuuQLGnxJF9B7yrE,7563
 wafer_core/rollouts/environments/repl.py,sha256=DyvyqEbBBytliqZ2-uAJqm-F7gdROG5-9LDFVyCo6lo,36042
@@ -501,7 +502,7 @@ wafer_core/rollouts/providers/openai_completions.py,sha256=3vUA74qjrxG-aOjyngtnZ
 wafer_core/rollouts/providers/openai_responses.py,sha256=xlyeI9h7aZEbpFY_8_zQ6IIYbMeNgcsEVRi92PBEAWc,32470
 wafer_core/rollouts/providers/sglang.py,sha256=kahYlrFG008D4jRA-c6mylBsTe-qTryKMhsjUbidduU,35525
 wafer_core/rollouts/templates/__init__.py,sha256=8qANHtoWZe9zpAWufkXlo8tQ07_Lw-RhX-lm2i-0ORQ,976
-wafer_core/rollouts/templates/base.py,sha256=FP6zx43ba8rx2nTx8dK1qZu-8sbmude2lBlM0xjSyGY,6906
+wafer_core/rollouts/templates/base.py,sha256=aBojLjssj2YaPjgCZMvHYWxtGa83Y4uqt1Cxoy1J3v0,7010
 wafer_core/rollouts/templates/loader.py,sha256=eNRWP2zMTsygpewBhAO3Co0vLb4o4SwJo74Jp1DWS-0,4726
 wafer_core/rollouts/tests/test_slice.py,sha256=jUJFkTUzY6TNrimFH5kkfOavn3J5Gg-hsFWUOzs30lk,9148
 wafer_core/rollouts/tools/__init__.py,sha256=nLEJrlhT1anqInh6u4CeX8CCrjUJpAEMp0jx-G0V6gs,298
@@ -584,11 +585,13 @@ wafer_core/sessions/dtypes.py,sha256=K6nOjvL6sjCGY7GTtdEygf1IZY_18R9YkHGqFyMd8wY
 wafer_core/sessions/hooks.py,sha256=A-txm6ufnRGQCdtP3vwh7oEOdlLN9Tv0XsjORMihuAI,4295
 wafer_core/targets/__init__.py,sha256=sHndC7AAOaHXlrmDXFLB53a5Y8DBjuyqS6nwsO2nj-Y,1728
 wafer_core/targets/digitalocean.py,sha256=cvoYpYjtSyy5t2lQAPi7ERruuuibronah_ivOiduAHQ,16550
-wafer_core/targets/runpod.py,sha256=bYTLVRaASrewJLIcZRtPMfMU2en1McE6W5w-Edo_OPQ,15785
-wafer_core/tools/__init__.py,sha256=zdWVdQU8aP66HVoI1AhoUHJ9UKG95rT5MFufD6zV2Q0,3212
+wafer_core/targets/runpod.py,sha256=LrVmNvA6qjzL5nbGSWvtw7CHrK6bDu7_o3vKIek00Tc,20286
+wafer_core/tools/__init__.py,sha256=deGQQlcdSD6zQx8JHizfSXgF5-EntdBOF_ngtob1-VU,3506
 wafer_core/tools/bash_tool.py,sha256=daoKOVGSgL0x9X_3l8Apd6-wFH4VMXMGJwVemw2FIfc,16828
 wafer_core/tools/glob_tool.py,sha256=9X5PdOjQJj7kiVNqqCZC0-1LmnE6wHx3Zc9zfMjtXdc,3533
 wafer_core/tools/grep_tool.py,sha256=cStyDz-J47oDLLZCL83yOvYo8Ijv4qu3D372JKT_ptM,4580
+wafer_core/tools/search_docs_tool.py,sha256=WY4hY83sseX8Fpxvw6DZxiG-F95F2t3-4PyfMD1Lpkg,6809
+wafer_core/tools/skill_tool.py,sha256=JXsT5hBTUH5U4tmzHEywU7eHHt5xCEF79tL2tsuk4-c,2067
 wafer_core/tools/wafer_tool.py,sha256=-dgPTHbWXq3I3wFj0mP7-lj5iZqGRoFvFf9IEEo3plQ,6345
 wafer_core/tools/write_kernel_tool.py,sha256=dJjhr-WBhVNe06hcJQVmBZTbS8mid64KF1MwlE2s2R4,21547
 wafer_core/tools/autotuner/BENCHMARKING.md,sha256=RkJ2wFhbDFXuMbw0mOW4pRqntT0UirptXwIxyrA1_KM,3825
@@ -613,7 +616,7 @@ wafer_core/tools/capture_tool/metrics.py,sha256=BFZNmdE-kh3LneYdWXTNZmlLuo-DCrP5
 wafer_core/tools/file_tools/__init__.py,sha256=2H7Rq5bijNQHGO4W6jjQAShkrcmdcHC0EQ8mBpgrApI,632
 wafer_core/tools/file_tools/edit_tool.py,sha256=Efx83pM1Ljb07cJmAGVhPX4YiPJICK70sZM6uCjRWB0,4109
 wafer_core/tools/file_tools/glob_tool.py,sha256=Av4LfC21fHXbnSsgh_9zDxlY9Qhb48aApaGos4j3B4g,3437
-wafer_core/tools/file_tools/grep_tool.py,sha256=FRIYeBfCcywHqkiT8OqVL8xOtMiZDOV6EhUU2LN5fqM,4444
+wafer_core/tools/file_tools/grep_tool.py,sha256=42eFj2pxBBrs5eg_GhyYJ-j2fNWkmGPvrEqXFmi5E10,5539
 wafer_core/tools/file_tools/read_tool.py,sha256=K0Hd8zwyL4Yva5YO9spXDfTRfXvfjqh9ztVrA8s1bJE,3961
 wafer_core/tools/file_tools/utils.py,sha256=HgaqYan2Pky4hTLX2L9d2Gj9oS325H7rFbJj-jryNtc,2576
 wafer_core/tools/file_tools/write_tool.py,sha256=X4N8y8wB-k9d5PcMRmZMRKIXlG9jHJiRdlEFFRLdZzs,2083
@@ -640,7 +643,7 @@ wafer_core/tools/tracelens_tools/tracelens_collective_tool.py,sha256=0E3FhfaA1N0
 wafer_core/tools/tracelens_tools/tracelens_compare_tool.py,sha256=99dUsB4wuYjxbh4X6Nsf2AtDMs94Uzy04tSemDOmKhg,4458
 wafer_core/tools/tracelens_tools/tracelens_report_tool.py,sha256=unuEx2zXaK42lA3qojS-WzFlBmIFrS75GHSgXUnDXGE,4720
 wafer_core/utils/__init__.py,sha256=oPHgkMkE7wS2lYKLlXrw4Ia5EHnpVcGHFfpWebIlVKs,354
-wafer_core/utils/backend.py,sha256=gNxv4jeOfZwrefTDhyp-UHzkWRyWRhmot76PRZkAwyk,8842
+wafer_core/utils/backend.py,sha256=zt5AX00OXSIstprvQ1_WNf_PQYphD2y53kOXWvg20RY,8986
 wafer_core/utils/code_validation.py,sha256=UqS4UVDxO-atdbn6i7JygX6IFPITvT56zZn1t-ZNuM8,4692
 wafer_core/utils/environment_serialization.py,sha256=cVDkapx0JC60CekazgirPEMAeGZhbLdX1WMIkFvId60,5047
 wafer_core/utils/event_streaming.py,sha256=Sg3-hI043Ofc2b29Z3DWrKgu4HkfJoIqhhbfGRJv70Q,2260
@@ -659,18 +662,18 @@ wafer_core/utils/kernel_utils/evaluate.py,sha256=1kxFNMl9VCXfKfk_BIiuA_zFfvDB1sl
 wafer_core/utils/kernel_utils/gpu_validation.py,sha256=LRiDjW_xAK4fXf1Vw1aYHG54B1W0J6b5L0K6PXzM2tI,3759
 wafer_core/utils/kernel_utils/reference_cache.py,sha256=4IQ2gND1StHULRO7geyAElEStbjQxwOeP6X09E5wCB0,11283
 wafer_core/utils/kernel_utils/results.py,sha256=QJGeah_41LSzxyYwGl9VxHPxTVAN2bLtk5bWdWLIpL4,6705
-wafer_core/utils/kernel_utils/static_checker.py,sha256=GWC7RZdwL4WqMLK0nO-wnZd-VmlMahCVKP6l9zVAW30,12909
+wafer_core/utils/kernel_utils/static_checker.py,sha256=XIQkzAOkGH5xtrOuZM4tNUqVJ0QRkYeJ7_8DosDOtkw,19886
 wafer_core/utils/kernel_utils/task.py,sha256=XcmKxKUWh5It6nX3zGqj77tWgA32uPfQMqNOqyD5T48,2682
 wafer_core/utils/kernel_utils/utils.py,sha256=uDZoJDxh07hJeLNlPdKN2vgB15pqIr1LbXf0YIBHU4E,43056
 wafer_core/utils/kernel_utils/targets/__init__.py,sha256=4NwRLsuJ__S4xKAfda4Ag82C5MQ3Qio-4xA5S-mQGlU,2067
-wafer_core/utils/kernel_utils/targets/config.py,sha256=eGpPOKlTxcS16Fs1-L9r9Cf_CDTG2-52bsG7LJllFw4,18900
+wafer_core/utils/kernel_utils/targets/config.py,sha256=sNXyYTZ9rL9OET4xqbHZ0d4b8ChzST1yUKvNOv8JSQs,19933
 wafer_core/utils/kernel_utils/targets/execution.py,sha256=bZuNXCo0sIdD6hFhetLPrtDC-zMSiIsAx_aml49VVL0,15033
 wafer_core/utils/kernel_utils/targets/selection.py,sha256=5I_RG_7cfhq7uaeR28meC2EeNNKssFsK-Tc3QFG6Ze0,3590
 wafer_core/utils/modal_execution/__init__.py,sha256=jkVqYOLzCT5K73N9Od0UIUsx-99A0m6bpDrxfyXxQZ8,945
-wafer_core/utils/modal_execution/modal_app.py,sha256=ibBllC59R9bP9w4QweFocVGeSX5kDmUuDl39PPYgCpE,11796
+wafer_core/utils/modal_execution/modal_app.py,sha256=VfS2cX8gHtnlPXemmMcEwDPeQdhiv2tly3CifOyh9f4,11455
 wafer_core/utils/modal_execution/modal_config.py,sha256=7cGX9TGqilQ3qxI3OFGXV5orjtyRU-PEDOJ4vP2oxno,4421
 wafer_core/utils/modal_execution/modal_execution.py,sha256=gChjnV6jqA3A7IRP3DfvV5cSfm_MN0X4f7JZufXgdZE,24594
 wafer_core/utils/modal_execution/test_modal.py,sha256=_jqou_hrLs1Daf1590Pnb0a_lXMMa2rczAPpW9HpoNQ,8153
-wafer_core-0.1.21.dist-info/METADATA,sha256=AmuwdcVbGfncJlEZCzq2gWq2LYhoi8rJMpHGe8OIbSo,1420
-wafer_core-0.1.21.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
-wafer_core-0.1.21.dist-info/RECORD,,
+wafer_core-0.1.23.dist-info/METADATA,sha256=HnIqBmqEQ6t_dc54Rnyg_Wyy-HKuAr3XTmsEoJkjJLo,1420
+wafer_core-0.1.23.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
+wafer_core-0.1.23.dist-info/RECORD,,