Add Gluon FA Blackwell for performance measurement (#268)

Summary:
Copy python/tutorials/gluon/01-attention-forward.py from OSS Version from PR7238

Pull Request resolved: #268

Reviewed By: xuzhao9

Differential Revision: D77901791

Pulled By: manman-ren

fbshipit-source-id: 4626816698fd55efc999f498660e02417b3b6b11

Author: manman-ren

test/test_gpu/skip_tests_h100_pytorch.yaml

@@ -12,6 +12,8 @@ flash_attention:
   - triton_tutorial_flash_v2_ws
   - triton_tutorial_flash_v2_tma_ws
   - triton_tutorial_flash_v2_tma_ws_persistent
+blackwell_attentions:
+  - gluon_blackwell_fwd
 # the two requires full fbgemm instead of genai flavor
 fp8_gemm_rowwise:
 fp8_gemm_rowwise_grouped:

test/test_gpu/skip_tests_h100_triton_main.yaml

@@ -12,6 +12,8 @@ flash_attention:
   - triton_tutorial_flash_v2_ws
   - triton_tutorial_flash_v2_tma_ws
   - triton_tutorial_flash_v2_tma_ws_persistent
+blackwell_attentions:
+  - gluon_blackwell_fwd
 # the two requires full fbgemm instead of genai flavor
 fp8_gemm_rowwise:
 fp8_gemm_rowwise_grouped:

tritonbench/kernels/attention_utils.py

@@ -16,6 +16,7 @@
 PEEL_LAST = os.getenv("PEEL_LAST_ITER")
 WITH_TMA = os.getenv("WITH_TMA")
 HAS_EXPLICIT_WS = os.getenv("ENABLE_EXPLICIT_WS")
+SUPPORT_GLUON = os.getenv("WITH_GLUON")
 
 
 class TmaAutoTuneHelper: