fix: resolve FP8 benchmark, NCCL parsing, and report None-value bugs

- benchmark.py: FP8 dtype now uses torch._scaled_mm() with scale tensors
  instead of torch.matmul() which does not support float8_e4m3fn on Hopper;
  fixes "addmm_cuda not implemented" error and enables FP8 TFLOPS measurement

- nccl_test.py: fix two bugs causing all-zero bandwidth results
  1. buffer size changed from -b 8 (8 bytes) to -b 8M -e 8G for meaningful load
  2. column parser corrected: parts[2] is dtype string not time value;
     now reads time=parts[5], algbw=parts[6], busbw=parts[7] per nccl-tests format

- report.py: replace .get(key, 0) with .get(key) or 0 at all bandwidth/stress
  fields to handle None values stored in result dicts (dict.get with default
  does not override an explicitly stored None)

- .gitignore: exclude .claude/settings.local.json

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

.gitignore

@@ -14,3 +14,4 @@ reports/
 .venv/
 venv/
 .qoder/*
+.claude/settings.local.json

modules/benchmark.py

@@ -349,15 +349,29 @@ class Benchmark:
 
                     if dtype_name == "fp8":
                         a = torch.randn(M, K, device="cuda", dtype=torch.float32).to(torch.float8_e4m3fn)
-                        b = torch.randn(K, N, device="cuda", dtype=torch.float32).to(torch.float8_e4m3fn)
+                        b = torch.randn(N, K, device="cuda", dtype=torch.float32).to(torch.float8_e4m3fn)
+                        scale_a = torch.tensor(1.0, device="cuda")
+                        scale_b = torch.tensor(1.0, device="cuda")
+                        def _fp8_mm():
+                            return torch._scaled_mm(a, b.T, scale_a=scale_a, scale_b=scale_b, out_dtype=torch.bfloat16)
                     else:
                         a = torch.randn(M, K, device="cuda", dtype=dtype_val)
                         b = torch.randn(K, N, device="cuda", dtype=dtype_val)
 
+                    if dtype_name == "fp8":
+                        for _ in range(warmup):
+                            _fp8_mm()
+                        torch.cuda.synchronize()
+                        start_event = torch.cuda.Event(enable_timing=True)
+                        end_event = torch.cuda.Event(enable_timing=True)
+                        start_event.record()
+                        for _ in range(iterations):
+                            c = _fp8_mm()
+                        end_event.record()
+                    else:
                         for _ in range(warmup):
                             torch.matmul(a, b)
                         torch.cuda.synchronize()
-
                         start_event = torch.cuda.Event(enable_timing=True)
                         end_event = torch.cuda.Event(enable_timing=True)
                         start_event.record()

modules/nccl_test.py

@@ -150,8 +150,8 @@ class NCCLTest:
 
         cmd = [
             binary,
-            "-b", "8",
+            "-b", "8M",
-            "-e", "256M",
+            "-e", "8G",
             "-f", "2",
             "-g", str(gpu_count),
             "-w", "5",
@@ -239,12 +239,15 @@ class NCCLTest:
             if not line or line.startswith("#"):
                 continue
             parts = line.split()
-            if len(parts) >= 7:
+            # nccl-tests data lines: size count type redop root time algbw busbw #wrong [time algbw busbw #wrong]
+            if len(parts) >= 9:
                 try:
                     size = int(parts[0])
-                    algbw = float(parts[-3]) if len(parts) >= 3 else 0
+                    # parts[2] is dtype string ('float'/'int32'/etc.), not a number
-                    busbw = float(parts[-2]) if len(parts) >= 2 else 0
+                    # out-of-place columns: time=parts[5], algbw=parts[6], busbw=parts[7]
-                    time_us = float(parts[2]) if len(parts) >= 3 else 0
+                    time_us = float(parts[5])
+                    algbw = float(parts[6])
+                    busbw = float(parts[7])
                     size_results.append({
                         "size": size,
                         "time_us": time_us,

modules/report.py

@@ -256,20 +256,20 @@ class ReportGenerator:
             lines.append(f"Source: {mem_data.get('source', 'unknown')}\n")
             lines.append("| Metric | Value | Peak | Efficiency |")
             lines.append("|--------|-------|------|------------|")
-            d2d = mem_data.get("d2d_bandwidth_gbps", 0)
+            d2d = mem_data.get("d2d_bandwidth_gbps") or 0
-            h2d = mem_data.get("h2d_bandwidth_gbps", 0)
+            h2d = mem_data.get("h2d_bandwidth_gbps") or 0
-            d2h = mem_data.get("d2h_bandwidth_gbps", 0)
+            d2h = mem_data.get("d2h_bandwidth_gbps") or 0
             # New format with per-metric peaks
-            h2d_peak = mem_data.get("h2d_peak_gbps", 0)
+            h2d_peak = mem_data.get("h2d_peak_gbps") or 0
-            d2h_peak = mem_data.get("d2h_peak_gbps", 0)
+            d2h_peak = mem_data.get("d2h_peak_gbps") or 0
-            d2d_peak = mem_data.get("d2d_peak_gbps", 0)
+            d2d_peak = mem_data.get("d2d_peak_gbps") or 0
-            h2d_eff = mem_data.get("h2d_efficiency_pct", 0)
+            h2d_eff = mem_data.get("h2d_efficiency_pct") or 0
-            d2h_eff = mem_data.get("d2h_efficiency_pct", 0)
+            d2h_eff = mem_data.get("d2h_efficiency_pct") or 0
-            d2d_eff = mem_data.get("d2d_efficiency_pct", 0)
+            d2d_eff = mem_data.get("d2d_efficiency_pct") or 0
             # Fallback for old format
             if not d2d_peak:
-                d2d_peak = mem_data.get("peak_bandwidth_gbps", 0)
+                d2d_peak = mem_data.get("peak_bandwidth_gbps") or 0
-                d2d_eff = mem_data.get("efficiency_pct", 0)
+                d2d_eff = mem_data.get("efficiency_pct") or 0
             lines.append(f"| H2D (PCIe) | {h2d:.1f} GB/s | {h2d_peak:.0f} GB/s | {h2d_eff:.1f}% |")
             lines.append(f"| D2H (PCIe) | {d2h:.1f} GB/s | {d2h_peak:.0f} GB/s | {d2h_eff:.1f}% |")
             lines.append(f"| D2D (NVLink) | {d2d:.1f} GB/s | {d2d_peak:.0f} GB/s | {d2d_eff:.1f}% |")
@@ -329,8 +329,8 @@ class ReportGenerator:
         if stress and not stress.get("error"):
             lines.append("## Stress Test\n")
             passed = stress.get("passed", False)
-            duration = stress.get("duration_sec", 0)
+            duration = stress.get("duration_sec") or 0
-            elapsed = stress.get("elapsed_sec", 0)
+            elapsed = stress.get("elapsed_sec") or 0
             source = stress.get("source", "unknown")
             lines.append(f"- **Source:** {source}")
             lines.append(f"- **Duration:** {elapsed:.0f}s (requested {duration}s)")
@@ -432,7 +432,7 @@ class ReportGenerator:
             if mem.get("error"):
                 items.append(("Memory Bandwidth", f"ERROR: {mem['error']}"))
             else:
-                eff = mem.get("efficiency_pct", 0)
+                eff = mem.get("efficiency_pct") or 0
                 verdict = "PASS" if eff >= 80 else ("WARN" if eff >= 50 else "FAIL")
                 items.append(("Memory Bandwidth", f"{verdict} ({eff:.1f}%)"))