misc: fix some B200 GEMM bench (#1883)

<!-- .github/pull_request_template.md -->

## 📌 Description
Before it couldn't run due to missing layout
<img width="1047" height="187" alt="image"
src="https://github.com/user-attachments/assets/c1995f60-d4f6-4ed6-8a72-97f24f7e8ec5"
/>

### After
<img width="894" height="356" alt="image"
src="https://github.com/user-attachments/assets/72f12a21-99f1-4374-8453-2965fe716856"
/>
<img width="891" height="227" alt="image"
src="https://github.com/user-attachments/assets/b7911d02-5914-4eb0-b69a-acd4d6c1595e"
/>


## 🔍 Related Issues

<!-- Link any related issues here -->

## 🚀 Pull Request Checklist

Thank you for contributing to FlashInfer! Before we review your pull
request, please make sure the following items are complete.

### ✅ Pre-commit Checks

- [ ] I have installed `pre-commit` by running `pip install pre-commit`
(or used your preferred method).
- [ ] I have installed the hooks with `pre-commit install`.
- [ ] I have run the hooks manually with `pre-commit run --all-files`
and fixed any reported issues.

> If you are unsure about how to set up `pre-commit`, see [the
pre-commit documentation](https://pre-commit.com/).

## 🧪 Tests

- [ ] Tests have been added or updated as needed.
- [ ] All tests are passing (`unittest`, etc.).

## Reviewer Notes

<!-- Optional: anything you'd like reviewers to focus on, concerns, etc.
-->

Author: Edenzzzz

benchmarks/bench_groupwise_gemm_fp8_blackwell.py

@@ -167,10 +167,12 @@ def bench_groupwise_gemm_fp8_blackwell(m, n, k, in_dtype, out_dtype):
     b_scale = torch.rand((k // 128, n // 128), dtype=torch.float32, device="cuda")
 
     out = torch.empty((m, n), dtype=out_dtype, device="cuda")
-    gemm_fp8_nt_groupwise(a, b, a_scale, b_scale, out=out)
+    gemm_fp8_nt_groupwise(a, b, a_scale, b_scale, out=out, scale_major_mode="MN")
 
     measurements = bench_gpu_time(
-        lambda: gemm_fp8_nt_groupwise(a, b, a_scale, b_scale, out=out)
+        lambda: gemm_fp8_nt_groupwise(
+            a, b, a_scale, b_scale, out=out, scale_major_mode="MN"
+        )
     )
     ms = np.median(measurements)
     tflops_per_second = 2 * m * n * k * 1e-9 / ms

benchmarks/bench_tgv_gemm.py

@@ -65,7 +65,7 @@ def test_tgv_gemm_bf16_sm100_perf():
 
     for m, n, k, has_bias, description in test_cases:
         print(f"\n--- {description}: M={m}, N={n}, K={k}, has_bias={has_bias} ---")
-
+        flops = m * n * k * 2 / 1e12
         # Create tensors
         A = torch.randn(m, k, device="cuda", dtype=torch.bfloat16)
         B = torch.randn(n, k, device="cuda", dtype=torch.bfloat16).t()
@@ -99,7 +99,9 @@ def test_tgv_gemm_bf16_sm100_perf():
         torch.cuda.synchronize()
         end_time = time.time()
         cublas_avg_time = (end_time - start_time) / 100
-        print(f"CUBLAS average time: {cublas_avg_time * 1000:.6f} ms")
+        print(
+            f"CUBLAS average time: {cublas_avg_time * 1000:.6f} ms, {flops / cublas_avg_time:.3f} TFLOPS"
+        )
 
         # Warmup
         with autotune(tune_mode=True):
@@ -128,7 +130,7 @@ def test_tgv_gemm_bf16_sm100_perf():
 
         tgv_avg_time = (end_time - start_time) / 100
         print(
-            f"TGV average time: {tgv_avg_time * 1000:.6f} ms, speedup: {cublas_avg_time / tgv_avg_time:.2f}x"
+            f"TGV average time: {tgv_avg_time * 1000:.6f} ms, {flops / tgv_avg_time:.3f} TFLOPS, speedup: {cublas_avg_time / tgv_avg_time:.2f}x"
         )
 
         # Test with PDL
@@ -151,7 +153,7 @@ def test_tgv_gemm_bf16_sm100_perf():
 
         pdl_avg_time = (end_time - start_time) / 100
         print(
-            f"PDL average time: {pdl_avg_time * 1000:.6f} ms, speedup: {cublas_avg_time / pdl_avg_time:.2f}x"
+            f"PDL average time: {pdl_avg_time * 1000:.6f} ms, {flops / pdl_avg_time:.3f} TFLOPS, speedup: {cublas_avg_time / pdl_avg_time:.2f}x"
         )
 
         # Store results for CSV