wafer-cli 0.2.32__py3-none-any.whl → 0.2.33__py3-none-any.whl

wafer/templates/aiter_optimize.py

@@ -0,0 +1,59 @@
+"""Template for optimizing AMD aiter operators.
+
+Usage:
+    wafer agent -t aiter-optimize --args op=gemm_a8w8 --args target=mi300x "Optimize this operator"
+    wafer agent -t aiter-optimize --args op=mha --args target=runpod-mi300x-rocm7 "Improve MHA performance"
+"""
+
+try:
+    from wafer.agent_defaults import (
+        AITER_BASH_ALLOWLIST,
+        AITER_ENABLED_TOOLS,
+        AITER_SYSTEM_PROMPT,
+    )
+except ImportError:
+    # Fallback for when wafer-cli package isn't installed
+    AITER_ENABLED_TOOLS = ["read", "write", "edit", "glob", "grep", "bash"]
+    AITER_BASH_ALLOWLIST = [
+        "ls", "cat", "head", "tail", "wc", "find", "grep", "rg", "pwd", "tree",
+        "which", "diff", "sort", "mkdir", "cp", "mv", "git diff", "git status",
+        "git log", "hipcc", "g++", "gcc", "clang", "python", "python3", "pip",
+        "pytest", "./", "wafer evaluate aiter", "wafer amd rocprof-compute",
+        "wafer amd rocprof-sdk", "wafer amd rocprof-systems", "wafer amd isa",
+        "wafer agent -t ask-docs", "timeout",
+    ]
+    AITER_SYSTEM_PROMPT = "You are a GPU kernel optimization expert for AMD MI300X and aiter."
+
+try:
+    from wafer_core.rollouts.templates import TemplateConfig
+except ImportError:
+    from rollouts.templates import TemplateConfig
+
+# Format system prompt with template variables ($op, $target become {op}, {target})
+# The template loader will substitute these at runtime
+_SYSTEM_PROMPT = AITER_SYSTEM_PROMPT.replace("{op}", "$op").replace("{target_flag}", "--target $target")
+
+template = TemplateConfig(
+    # Identity
+    name="aiter-optimize",
+    description="Optimize AMD aiter operators for better performance on MI300X",
+    # System prompt - uses shared prompt from agent_defaults
+    system_prompt=_SYSTEM_PROMPT,
+    # Tools - full coding environment
+    tools=AITER_ENABLED_TOOLS,
+    bash_allowlist=AITER_BASH_ALLOWLIST,
+    # Network access required for wafer evaluate (connects to remote GPU)
+    allow_network=True,
+    # Model config - use thinking for optimization analysis
+    model="anthropic/claude-sonnet-4-5-20250929",
+    max_tokens=16384,
+    thinking=True,
+    thinking_budget=10000,
+    # Multi-turn for iterative optimization
+    single_turn=False,
+    # Template variables
+    defaults={
+        "op": "gemm_a8w8",
+        "target": "mi300x",  # Required - user must specify their target
+    },
+)

wafer/templates/optimize_kernel.py

@@ -35,8 +35,7 @@ Strategy:
 Commands:
 - `wafer evaluate --impl <file> --reference <ref> --test-cases <tests>` - Run evaluation
 - `wafer evaluate --impl <file> --reference <ref> --test-cases <tests> --profile` - With NCU profiling
-- `wafer workspaces exec -- <command>` - Run arbitrary commands on remote GPU
-- `wafer targets exec <target> -- <command>` - Run commands on a configured target via SSH
+- `wafer remote-run "<command>"` - Run arbitrary commands on remote GPU
 
 Output:
 - Summary of optimizations applied
@@ -49,8 +48,7 @@ IMPORTANT: Always verify correctness with wafer evaluate before claiming success
     tools=["read", "write", "edit", "glob", "grep", "bash"],
     bash_allowlist=[
         "wafer evaluate",
-        "wafer workspaces exec",
-        "wafer targets exec",
+        "wafer remote-run",
         "wafer nvidia ncu",
         "wafer nvidia nsys",
         "wafer nvidia perfetto",

wafer/templates/optimize_kernelbench.py

@@ -1,4 +1,4 @@
-"""Template for KernelBench optimization.
+"""Template for KernelBench optimization - matches eval system prompt.
 
 Usage:
     # Run on a specific problem
@@ -26,18 +26,12 @@ try:
 except ImportError:
     from rollouts.templates import TemplateConfig
 
-from wafer.agent_defaults import ENABLED_TOOLS, KERNELBENCH_BASH_ALLOWLIST
-
-# Task-specific instructions only — must stay in sync with the eval's SYSTEM_PROMPT
-# in research/evals/optimize_kernelbench_eval/.../base_config.py.
-# Run test_eval_cli_parity.py to verify.
-# Wafer CLI command docs are auto-generated from --help text and composed
-# at runtime by wevin_cli.py (see wafer.cli_instructions.build_cli_instructions).
-# TODO: Consider having both eval and template import SYSTEM_PROMPT from a shared
-# module so there's only one copy to maintain.
+# System prompt matches optimize_kernelbench_eval/base_config.py SYSTEM_PROMPT
 SYSTEM_PROMPT = """\
 You are a GPU kernel optimization expert. Your task is to write optimized GPU kernels that are correct and faster than the PyTorch baseline.
 
+IMPORTANT: You do NOT have a local GPU. You MUST use `wafer evaluate kernelbench` to test kernels on remote GPU hardware.
+
 ## Kernel Format (KernelBench)
 
 The reference file contains a PyTorch `Model` class. You must write a `ModelNew` class that:
@@ -49,14 +43,49 @@ The reference file also provides:
 - `get_inputs()` - generates test inputs for forward()
 - `get_init_inputs()` - generates constructor arguments
 
+## Available Tools
+
+- read(file_path): Read source files
+- write(file_path, content): Write your optimized kernel
+- glob(pattern): Find files by pattern
+- grep(pattern): Search code
+- bash(command): Run shell commands including wafer CLI
+
 ## Workflow
 
 1. Read the reference problem file to understand what `Model` does
 2. Analyze the computation and identify optimization opportunities
 3. Write an optimized `ModelNew` class with custom $backend_upper kernels using `__global__` kernel definitions and `torch.utils.cpp_extension.load_inline`
-4. Test with: `wafer evaluate kernelbench $target_flag --backend $backend --impl optimized.py --reference <problem.py> --benchmark`
+4. Test with: `wafer evaluate kernelbench $target_flag --backend $backend --impl <your_file.py> --reference <problem.py> --benchmark`
 5. Iterate based on feedback until correct and fast
 
+## Example Command
+
+```bash
+wafer evaluate kernelbench \\
+    $target_flag \\
+    --backend $backend \\
+    --impl optimized_kernel.py \\
+    --reference $reference \\
+    --benchmark
+```
+
+## Profiling Tools (USE THESE!)
+
+When your kernel is slower than expected, use profiling to understand WHY:
+
+- `wafer rocprof profile --impl <file> --reference <ref>` - AMD GPU profiling
+- `wafer nvidia ncu --impl <file> --reference <ref>` - NVIDIA NCU profiling
+
+## CRITICAL: Reactive Debugging
+
+After EVERY `wafer evaluate` call:
+1. Check the speedup result
+2. If speedup < 1.0x (slowdown), STOP and analyze:
+   - Run profiling to identify the bottleneck
+   - Ask: "Why is this slow?" before trying another approach
+3. Don't just try random optimizations - understand the root cause
+
 Your kernel MUST:
 - Pass correctness tests (outputs match reference within tolerance)
 - Achieve speedup > 1.0x over PyTorch baseline
@@ -67,16 +96,32 @@ You MUST run `wafer evaluate kernelbench` to verify your kernel. Your score depe
 template = TemplateConfig(
     # Identity
     name="optimize-kernelbench",
-    description="Optimize KernelBench problems",
-    # System prompt (task-specific; CLI docs appended at runtime)
+    description="Optimize KernelBench problems (matches eval system prompt)",
+    # System prompt
     system_prompt=SYSTEM_PROMPT,
     # Tools
-    tools=ENABLED_TOOLS,
-    bash_allowlist=KERNELBENCH_BASH_ALLOWLIST,
-    # Model config
+    tools=["read", "write", "edit", "glob", "grep", "bash"],
+    bash_allowlist=[
+        "wafer evaluate",
+        "wafer nvidia ncu",
+        "wafer nvidia nsys",
+        "wafer rocprof",
+        "wafer compiler-analyze",
+        "python",
+        "python3",
+        "timeout",
+        "ls",
+        "cat",
+        "head",
+        "tail",
+        "wc",
+        "pwd",
+        "which",
+    ],
+    # Model config - match eval settings
     model="anthropic/claude-opus-4-5-20251101",
     max_tokens=8192,
-    # No thinking by default, can override with --thinking
+    # No thinking by default (match eval), can override with --thinking
     thinking=False,
     # Multi-turn for iterative optimization
     single_turn=False,

wafer/templates/optimize_vllm.py

@@ -0,0 +1,156 @@
+"""Template for vLLM kernel optimization.
+
+Usage:
+    # Optimize fused_moe kernel
+    wafer agent -t optimize-vllm \
+        --args vllm_dir=/path/to/vllm \
+        --args op=fused_moe \
+        --args target=my-gpu-server \
+        "Optimize the fused MoE kernel for better throughput"
+
+    # With custom test and benchmark commands
+    wafer agent -t optimize-vllm \
+        --args vllm_dir=./vllm \
+        --args op=paged_attention \
+        --args test_cmd="pytest tests/kernels/attention/test_attention.py -v" \
+        --args bench_cmd="python benchmarks/kernels/benchmark_paged_attention.py" \
+        --json
+
+Variables:
+    - vllm_dir: Path to vLLM repository (required)
+    - op: Target op to optimize (required, e.g., fused_moe, paged_attention)
+    - target: Target name (default: uses default target)
+    - pool: Target pool name (alternative to target)
+    - test_cmd: Pytest command for correctness (auto-generated from op if not provided)
+    - bench_cmd: Kernel microbenchmark command (auto-generated from op if not provided)
+"""
+
+try:
+    from wafer_core.rollouts.templates import TemplateConfig
+except ImportError:
+    from rollouts.templates import TemplateConfig
+
+from wafer.agent_defaults import VLLM_BASH_ALLOWLIST, VLLM_ENABLED_TOOLS
+
+# Default test commands per op (from vLLM's test structure)
+DEFAULT_TEST_CMDS = {
+    "fused_moe": "pytest tests/kernels/moe/test_moe.py -v",
+    "paged_attention": "pytest tests/kernels/attention/test_attention.py -v",
+    "flash_attn": "pytest tests/kernels/attention/test_flash_attn.py -v",
+    "flashinfer": "pytest tests/kernels/attention/test_flashinfer.py -v",
+    "rms_norm": "pytest tests/kernels/core/test_layernorm.py -v -k rms",
+    "layernorm": "pytest tests/kernels/core/test_layernorm.py -v",
+    "rotary_embedding": "pytest tests/kernels/core/test_rotary_embedding.py -v",
+    "activation": "pytest tests/kernels/core/test_activation.py -v",
+    "fused_topk": "pytest tests/kernels/moe/test_fused_topk.py -v",
+    "fp8_quant": "pytest tests/kernels/quantization/test_fp8_quant.py -v",
+    "int8_quant": "pytest tests/kernels/quantization/test_int8_quant.py -v",
+}
+
+# Default benchmark commands per op.
+# Uses pytest with --durations to measure kernel execution time.
+# vLLM v0.15+ kernel benchmarks require config context, so pytest
+# (which sets up fixtures) is the reliable path.
+DEFAULT_BENCH_CMDS = {
+    "fused_moe": "pytest tests/kernels/moe/test_moe.py --timeout=300 --durations=0 -q",
+    "paged_attention": "pytest tests/kernels/attention/test_attention.py --timeout=300 --durations=0 -q",
+    "rms_norm": "pytest tests/kernels/core/test_layernorm.py -k rms --timeout=120 --durations=0 -q",
+    "layernorm": "pytest tests/kernels/core/test_layernorm.py --timeout=120 --durations=0 -q",
+    "rotary_embedding": "pytest tests/kernels/core/test_rotary_embedding.py --timeout=120 --durations=0 -q",
+    "activation": "pytest tests/kernels/core/test_activation.py --timeout=120 --durations=0 -q",
+    "fused_topk": "pytest tests/kernels/moe/test_fused_topk.py --timeout=120 --durations=0 -q",
+    "fp8_quant": "pytest tests/kernels/quantization/test_fp8_quant.py --timeout=120 --durations=0 -q",
+    "int8_quant": "pytest tests/kernels/quantization/test_int8_quant.py --timeout=120 --durations=0 -q",
+}
+
+SYSTEM_PROMPT = """\
+You are a GPU kernel optimization expert. Your task is to improve the performance
+of a specific vLLM kernel while maintaining correctness.
+
+## Target
+
+You are optimizing the `$op` kernel in vLLM.
+- vLLM directory: `$vllm_dir`
+- Correctness test: `$test_cmd`
+- Benchmark: `$bench_cmd`
+
+## Workflow
+
+1. **Understand the kernel**: Read the kernel implementation in `$vllm_dir`
+   - For MoE: `vllm/model_executor/layers/fused_moe/`
+   - For attention: `vllm/attention/backends/`
+   - For normalization: `vllm/_custom_ops.py` or specific layer files
+   - For quantization: `vllm/_custom_ops.py`
+
+2. **Run baseline benchmark**: Establish baseline performance
+   ```bash
+   cd $vllm_dir && $bench_cmd
+   ```
+
+3. **Analyze and optimize**: Identify optimization opportunities
+   - Memory access patterns (coalescing, shared memory usage)
+   - Occupancy and register pressure
+   - Algorithm improvements
+   - Hardware-specific optimizations (tensor cores, etc.)
+
+4. **Modify the kernel**: Make your changes to improve performance
+
+5. **Validate correctness**: Run the test suite
+   ```bash
+   cd $vllm_dir && $test_cmd
+   ```
+
+6. **Measure improvement**: Run benchmark again and compare
+
+7. **Iterate**: If correctness fails or performance regresses, adjust and retry
+
+## Evaluation
+
+Use the wafer evaluate command to run both correctness and benchmark:
+```bash
+wafer evaluate vllm --vllm-dir $vllm_dir --op $op \\
+    --test-cmd "$test_cmd" \\
+    --bench-cmd "$bench_cmd" \\
+    $target_flag --json
+```
+
+## Constraints
+
+- The correctness test MUST pass after your changes
+- Focus on the specific kernel identified (`$op`)
+- Document your changes and reasoning
+- Your score depends on actual measured throughput improvement
+
+## Key Metrics
+
+- **time_us**: kernel execution time in microseconds (lower is better)
+- **tflops**: teraflops achieved (higher is better)
+- **bandwidth_gbps**: memory bandwidth in GB/s (higher is better)"""
+
+template = TemplateConfig(
+    # Identity
+    name="optimize-vllm",
+    description="Optimize vLLM kernels for better inference performance",
+    # System prompt (task-specific; CLI docs appended at runtime)
+    system_prompt=SYSTEM_PROMPT,
+    # Tools
+    tools=VLLM_ENABLED_TOOLS,
+    bash_allowlist=VLLM_BASH_ALLOWLIST,
+    # Model config
+    model="anthropic/claude-opus-4-5-20251101",
+    max_tokens=8192,
+    # No thinking by default, can override with --thinking
+    thinking=False,
+    # Multi-turn for iterative optimization
+    single_turn=False,
+    # Template variables
+    defaults={
+        "vllm_dir": "./vllm",
+        "op": "fused_moe",
+        "target": "",
+        "pool": "",
+        "test_cmd": "",  # Auto-filled from DEFAULT_TEST_CMDS[op] if empty
+        "bench_cmd": "",  # Auto-filled from DEFAULT_BENCH_CMDS[op] if empty
+        "target_flag": "",  # Auto-computed: --target X or --pool Y
+    },
+)

wafer/trace_compare.py

@@ -6,22 +6,19 @@ All core logic is in wafer_core.lib.trace_compare.
 
 import sys
 from pathlib import Path
-from typing import Any
 
 import typer
 
-import json
-import sys
-
 from wafer_core.lib.trace_compare import (
-    analyze_trace_pair,
+    analyze_fusion_differences,
+    analyze_traces,
     format_csv,
+    format_fusion_csv,
+    format_fusion_json,
+    format_fusion_text,
     format_json,
     format_text,
-    ArchitectureType,
-    detect_architecture,
 )
-from wafer_core.lib.trace_compare.loader import StreamingMetadata
 
 
 def compare_traces(
@@ -33,7 +30,6 @@ def compare_traces(
     show_layers: bool = False,
     show_all: bool = False,
     show_stack_traces: bool = False,
-    recommendations: bool = False,
 ) -> None:
     """Compare two GPU traces and generate performance report.
 
@@ -56,60 +52,21 @@ def compare_traces(
         typer.secho(f"❌ File not found: {trace2}", fg=typer.colors.RED, err=True)
         raise typer.Exit(1)
 
-    # Progress callback for JSON format (emits NDJSON to stdout)
-    def progress_callback(stage: str, fraction: float) -> None:
-        if output_format == 'json':
-            progress_msg = json.dumps({"type": "progress", "stage": stage, "fraction": fraction})
-            print(progress_msg, file=sys.stdout, flush=True)
-        elif output_format != 'json':
-            percent = int(fraction * 100)
-            typer.echo(f"📊 {stage}: {percent}%", err=True)
-
-    # Metadata callback for JSON format (emits NDJSON with early GPU info)
-    def metadata_callback(meta1: StreamingMetadata, meta2: StreamingMetadata) -> None:
-        if output_format == 'json':
-            metadata_msg = json.dumps({
-                "type": "metadata",
-                "trace1": {
-                    "platform": meta1.platform,
-                    "gpu": meta1.gpu_name,
-                    "file_size_mb": round(meta1.file_size_mb, 1),
-                },
-                "trace2": {
-                    "platform": meta2.platform,
-                    "gpu": meta2.gpu_name,
-                    "file_size_mb": round(meta2.file_size_mb, 1),
-                },
-            })
-            print(metadata_msg, file=sys.stdout, flush=True)
-        else:
-            typer.echo(f"📊 Trace 1: {meta1.platform} - {meta1.gpu_name} ({meta1.file_size_mb:.1f}MB)", err=True)
-            typer.echo(f"📊 Trace 2: {meta2.platform} - {meta2.gpu_name} ({meta2.file_size_mb:.1f}MB)", err=True)
-
-    # Analyze traces using unified API
+    # Analyze traces
+    # Only show progress messages for non-JSON formats (JSON needs clean stdout)
     if output_format != 'json':
         typer.echo("📊 Loading traces...")
 
+    # Determine how many stack traces to collect
+    max_stacks = 0 if (show_stack_traces and show_all) else (3 if show_stack_traces else 3)
+
     try:
-        result_obj = analyze_trace_pair(
+        results = analyze_traces(
             trace1,
             trace2,
-            phase=phase,
-            include_stacks=True,
-            on_progress=progress_callback,
-            on_metadata=metadata_callback,
+            phase_filter=phase,
+            max_stacks=max_stacks,
         )
-        
-        results = {
-            "metadata": result_obj.metadata,
-            "operations": result_obj.operations,
-            "layers": result_obj.layers,
-            "warnings": [{"code": w.code, "severity": w.severity, "message": w.message, "suggestion": w.suggestion} for w in result_obj.warnings],
-            "architecture": result_obj.architecture.value,
-            "layer_alignments": result_obj.layer_alignments,
-            "fusion_analysis": result_obj.fusion_analysis,
-            "same_kernel_analysis": result_obj.same_kernel_analysis,
-        }
     except ValueError as e:
         typer.secho(f"❌ {e}", fg=typer.colors.RED, err=True)
         raise typer.Exit(1)
@@ -117,26 +74,17 @@ def compare_traces(
         typer.secho(f"❌ Error analyzing traces: {e}", fg=typer.colors.RED, err=True)
         raise typer.Exit(1)
 
+    # Show loading confirmation
     if output_format != 'json':
         meta = results["metadata"]
+        # Determine which trace is AMD and which is NVIDIA
         if meta['trace1_platform'] == 'AMD':
             amd_gpu, nvidia_gpu = meta['trace1_gpu'], meta['trace2_gpu']
         else:
             amd_gpu, nvidia_gpu = meta['trace2_gpu'], meta['trace1_gpu']
         typer.echo(f"✅ Loaded: AMD ({amd_gpu}) vs NVIDIA ({nvidia_gpu})")
-        
-        # Display warnings
-        warnings = results.get("warnings", [])
-        if warnings:
-            typer.echo()
-            for warning in warnings:
-                icon = "❌" if warning["severity"] == "error" else "⚠️" if warning["severity"] == "warning" else "ℹ️"
-                typer.secho(f"{icon}  {warning['message']}", fg=typer.colors.YELLOW if warning["severity"] == "warning" else typer.colors.BLUE)
-                if warning.get("suggestion"):
-                    typer.secho(f"   Suggestion: {warning['suggestion']}", fg=typer.colors.BLUE)
     typer.echo()
 
-
     # Generate output based on format
     if output_format == "text":
         output_str = format_text(results, show_layers=show_layers, show_all=show_all, show_stack_traces=show_stack_traces)
@@ -160,23 +108,21 @@ def compare_traces(
         typer.echo(output_str)
 
 
-def compare_align(
+def compare_fusion(
     trace1: Path,
     trace2: Path,
     output: Path | None = None,
-    output_format: str = "json",
-    phase: str = "all",
-    layer: int | None = None,
+    format_type: str = "text",
+    min_group_size: int = 50,
 ) -> None:
-    """Align kernels at layer level for exact kernel-to-kernel comparison.
+    """Analyze kernel fusion differences between AMD and NVIDIA traces.
 
     Args:
         trace1: Path to first trace file (AMD or NVIDIA)
         trace2: Path to second trace file (AMD or NVIDIA)
         output: Optional output file path (default: stdout)
-        output_format: Output format ('json' only for now)
-        phase: Filter by phase ('all', 'prefill', or 'decode')
-        layer: Focus on specific layer number (optional)
+        format_type: Output format ('text', 'csv', or 'json')
+        min_group_size: Minimum correlation group size to analyze
     """
     # Validate files exist
     if not trace1.exists():
@@ -187,86 +133,49 @@ def compare_align(
         typer.secho(f"❌ File not found: {trace2}", fg=typer.colors.RED, err=True)
         raise typer.Exit(1)
 
-    # Progress callback for JSON format (emits NDJSON to stdout)
-    def progress_callback(stage: str, fraction: float) -> None:
-        if output_format == 'json':
-            progress_msg = json.dumps({"type": "progress", "stage": stage, "fraction": fraction})
-            print(progress_msg, file=sys.stdout, flush=True)
-        else:
-            percent = int(fraction * 100)
-            typer.echo(f"📊 {stage}: {percent}%", err=True)
-
-    # Metadata callback for JSON format
-    def metadata_callback(meta1: StreamingMetadata, meta2: StreamingMetadata) -> None:
-        if output_format == 'json':
-            metadata_msg = json.dumps({
-                "type": "metadata",
-                "trace1": {
-                    "platform": meta1.platform,
-                    "gpu": meta1.gpu_name,
-                    "file_size_mb": round(meta1.file_size_mb, 1),
-                },
-                "trace2": {
-                    "platform": meta2.platform,
-                    "gpu": meta2.gpu_name,
-                    "file_size_mb": round(meta2.file_size_mb, 1),
-                },
-            })
-            print(metadata_msg, file=sys.stdout, flush=True)
-        else:
-            typer.echo(f"📊 Trace 1: {meta1.platform} - {meta1.gpu_name} ({meta1.file_size_mb:.1f}MB)", err=True)
-            typer.echo(f"📊 Trace 2: {meta2.platform} - {meta2.gpu_name} ({meta2.file_size_mb:.1f}MB)", err=True)
-
-    # Analyze traces using unified API
-    if output_format != 'json':
+    # Analyze fusion
+    # Only show progress messages for non-JSON formats (JSON needs clean stdout)
+    if format_type != 'json':
         typer.echo("📊 Loading traces...")
-
     try:
-        result_obj = analyze_trace_pair(
+        results = analyze_fusion_differences(
             trace1,
             trace2,
-            phase=phase,
-            include_stacks=True,
-            on_progress=progress_callback,
-            on_metadata=metadata_callback,
+            min_group_size=min_group_size,
         )
-        
-        results = {
-            "metadata": result_obj.metadata,
-            "layer_alignments": result_obj.layer_alignments or [],
-            "fusion_analysis": result_obj.fusion_analysis or {},
-            "same_kernel_analysis": result_obj.same_kernel_analysis or {},
-            "operations": result_obj.operations,
-            "layers": result_obj.layers,
-            "warnings": [{"code": w.code, "severity": w.severity, "message": w.message, "suggestion": w.suggestion} for w in result_obj.warnings],
-            "architecture": result_obj.architecture.value,
-        }
-        
-        if layer is not None:
-            results["layer_alignments"] = [
-                la for la in results["layer_alignments"] if la.get("layer") == layer
-            ]
-    except ValueError as e:
-        typer.secho(f"❌ {e}", fg=typer.colors.RED, err=True)
-        raise typer.Exit(1)
     except Exception as e:
-        typer.secho(f"❌ Error analyzing traces: {e}", fg=typer.colors.RED, err=True)
+        typer.secho(
+            f"❌ Error analyzing traces: {e}", fg=typer.colors.RED, err=True
+        )
         import traceback
+
         traceback.print_exc()
         raise typer.Exit(1)
 
-    if output_format != 'json':
+    # Show loading confirmation
+    if format_type != 'json':
         meta = results["metadata"]
-        typer.echo(f"✅ Loaded: {meta.get('amd_gpu', 'Unknown')} vs {meta.get('nvidia_gpu', 'Unknown')}")
-        typer.echo(f"✅ Found {len(results['layer_alignments'])} layers")
+        # Note: fusion analyzer always uses trace1=AMD, trace2=NVIDIA
+        typer.echo(f"✅ Loaded: {meta['trace1_gpu']} vs {meta['trace2_gpu']}")
+        typer.echo(
+            f"Found {meta['trace1_correlation_groups']} trace1 groups and "
+            f"{meta['trace2_correlation_groups']} trace2 groups with ≥{min_group_size} kernels"
+        )
+        typer.echo(f"✅ Matched {meta['matched_groups']} correlation groups")
         typer.echo()
 
-    if output_format == "json":
-        output_str = format_json(results)
+    # Generate output
+    if format_type == "text":
+        output_str = format_fusion_text(results)
+    elif format_type == "csv":
+        output_str = format_fusion_csv(results)
+    elif format_type == "json":
+        output_str = format_fusion_json(results)
     else:
-        typer.secho(f"❌ Format {output_format} not yet supported for align command. Use 'json'.", fg=typer.colors.RED, err=True)
+        typer.secho(f"❌ Unknown format: {format_type}", fg=typer.colors.RED, err=True)
         raise typer.Exit(1)
 
+    # Write output
     if output:
         output.write_text(output_str)
         typer.secho(f"✅ Report saved to {output}", fg=typer.colors.GREEN)

wafer/wevin_cli.py

@@ -550,7 +550,7 @@ def main(  # noqa: PLR0913, PLR0915
     api_base, api_key, api_key_refresh = _get_wafer_auth(no_proxy=no_proxy)
     if not api_base or not api_key:
         print("Error: No API credentials found", file=sys.stderr)
-        print("  Run 'wafer auth login' or set ANTHROPIC_API_KEY", file=sys.stderr)
+        print("  Run 'wafer login' or set ANTHROPIC_API_KEY", file=sys.stderr)
         sys.exit(1)
 
     assert api_base is not None
@@ -573,17 +573,6 @@ def main(  # noqa: PLR0913, PLR0915
         tpl = _get_default_template()
         base_system_prompt = tpl.system_prompt
 
-    # Compose CLI instructions from --help text for allowed wafer commands
-    # TODO: The eval path doesn't have the skills layer below. If include_skills
-    # is ever enabled for optimize-kernelbench, the eval would need it too for parity.
-    # See test_eval_cli_parity.py for coverage notes.
-    if tpl.bash_allowlist:
-        from wafer.cli_instructions import build_cli_instructions
-
-        cli_instructions = build_cli_instructions(tpl.bash_allowlist)
-        if cli_instructions:
-            base_system_prompt = base_system_prompt + "\n\n" + cli_instructions
-
     # Append skill metadata if skills are enabled
     if tpl.include_skills:
         from wafer_core.rollouts.skills import discover_skills, format_skill_metadata_for_prompt