Skip to content

bench: Enable microbenchmarking on SM121 #3002

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
bkryu merged 3 commits into from
Apr 7, 2026
Merged

bench: Enable microbenchmarking on SM121 #3002

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
6548ae5
Add spark to microbenchmarks
bkryu Apr 6, 2026
951d414
CUTLASS MoE
bkryu Apr 7, 2026
0d2168a
Merge branch 'main' into bench_sm121
bkryu Apr 7, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
49 changes: 48 additions & 1 deletion benchmarks/routines/flashinfer_benchmark_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -298,6 +298,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["fa2", "fa2_tc", "auto", "cudnn", "trtllm-gen", "trtllm-native"],
"10.3": ["fa2", "fa2_tc", "auto", "cudnn", "trtllm-gen", "trtllm-native"],
"12.0": ["fa2", "fa2_tc", "auto", "cudnn", "trtllm-native"],
"12.1": ["fa2", "fa2_tc", "auto", "cudnn", "trtllm-native"],
},
"BatchPrefillWithPagedKVCacheWrapper": {
# NOTE: trtllm-native calls trtllm_batch_context_with_kv_cache
Expand All @@ -310,6 +311,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["fa2", "auto", "cudnn", "cudnn-native", "trtllm-gen", "trtllm-native"],
"10.3": ["fa2", "auto", "cudnn", "cudnn-native", "trtllm-gen", "trtllm-native"],
"12.0": ["fa2", "auto", "cudnn", "cudnn-native"],
"12.1": ["fa2", "auto", "cudnn", "cudnn-native"],
},
"BatchPrefillWithRaggedKVCacheWrapper": {
# NOTE: trtllm-native calls trtllm_ragged_attention_deepseek
Expand All @@ -322,6 +324,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["fa2", "cudnn", "cudnn-native", "cutlass", "trtllm-native"],
"10.3": ["fa2", "cudnn", "cudnn-native", "cutlass", "trtllm-native"],
"12.0": ["fa2", "cudnn", "cudnn-native"],
"12.1": ["fa2", "cudnn", "cudnn-native"],
},
"BatchMLAPagedAttentionWrapper": {
# NOTE: trtllm-native calls trtllm_batch_decode_with_kv_cache_mla
Expand All @@ -334,6 +337,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["fa2", "cutlass", "trtllm-native", "cute-dsl"],
"10.3": ["fa2", "cutlass", "trtllm-native"],
"12.0": ["fa2"],
"12.1": ["fa2"],
},
# GEMM
"gemm_fp8_nt_groupwise": {
Expand All @@ -345,6 +349,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cutlass"],
"10.3": ["cutlass"],
"12.0": [],
"12.1": [],
},
"group_gemm_fp8_nt_groupwise": {
"7.5": [],
Expand All @@ -355,6 +360,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cutlass"],
"10.3": ["cutlass"],
"12.0": [],
"12.1": [],
},
"bmm_fp8": {
"7.5": [],
Expand All @@ -365,6 +371,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cudnn", "cublas", "cutlass"],
"10.3": ["cudnn", "cublas", "cutlass"],
"12.0": ["cudnn", "cublas"],
"12.1": ["cudnn", "cublas"],
},
"bmm_mxfp8": {
"7.5": [],
Expand All @@ -375,6 +382,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cudnn"],
"10.3": ["cudnn"],
"12.0": [],
"12.1": [],
},
"mm_mxfp8": {
"7.5": [],
Expand All @@ -386,6 +394,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.3": ["cutlass", "cute-dsl", "trtllm"],
"11.0": ["cutlass"],
"12.0": [],
"12.1": [],
},
# Note: mm_fp4, mm_bf16, and bmm_bf16 use support checkers to filter backends, so they are not listed here
# MOE
Expand All @@ -398,6 +407,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["trtllm"],
"10.3": ["trtllm"],
"12.0": [],
"12.1": [],
},
"trtllm_fp8_block_scale_moe": {
"7.5": [],
Expand All @@ -408,6 +418,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["trtllm"],
"10.3": ["trtllm"],
"12.0": [],
"12.1": [],
},
"trtllm_fp8_per_tensor_scale_moe": {
"7.5": [],
Expand All @@ -418,6 +429,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["trtllm"],
"10.3": ["trtllm"],
"12.0": [],
"12.1": [],
},
"cutlass_fused_moe": {
"7.5": [],
Expand All @@ -427,7 +439,8 @@ def dtype_str_to_torch_dtype(dtype_str):
"9.0": [],
"10.0": ["cutlass"],
"10.3": ["cutlass"],
"12.0": [],
"12.0": ["cutlass"],
"12.1": ["cutlass"],
},
"cute_dsl_fp4_block_scale_moe": {
"7.5": [],
Expand All @@ -438,6 +451,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cute-dsl"],
"10.3": ["cute-dsl"],
"12.0": [],
"12.1": [],
},
# NORM
"rmsnorm": {
Expand All @@ -449,6 +463,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cuda"],
"10.3": ["cuda"],
"12.0": ["cuda"],
"12.1": ["cuda"],
},
"rmsnorm_quant": {
"7.5": ["cuda"],
Expand All @@ -459,6 +474,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cuda"],
"10.3": ["cuda"],
"12.0": ["cuda"],
"12.1": ["cuda"],
},
"fused_add_rmsnorm_quant": {
"7.5": ["cuda"],
Expand All @@ -469,6 +485,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cuda"],
"10.3": ["cuda"],
"12.0": ["cuda"],
"12.1": ["cuda"],
},
# NORM - FP4 Quantization (Blackwell SM100+ only, CuTe-DSL kernels)
"rmsnorm_fp4quant": {
Expand All @@ -480,6 +497,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cute-dsl"],
"10.3": ["cute-dsl"],
"12.0": ["cute-dsl"],
"12.1": ["cute-dsl"],
},
"add_rmsnorm_fp4quant": {
"7.5": [],
Expand All @@ -490,6 +508,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cute-dsl"],
"10.3": ["cute-dsl"],
"12.0": ["cute-dsl"],
"12.1": ["cute-dsl"],
},
# QUANTIZATION
"mxfp8_quantize": {
Expand All @@ -501,6 +520,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cuda", "cute-dsl"],
"10.3": ["cuda", "cute-dsl"],
"12.0": ["cuda", "cute-dsl"],
"12.1": ["cuda", "cute-dsl"],
},
"mxfp4_quantize": {
"7.5": [],
Expand All @@ -511,6 +531,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cuda", "cute-dsl"],
"10.3": ["cuda", "cute-dsl"],
"12.0": ["cuda", "cute-dsl"],
"12.1": ["cuda", "cute-dsl"],
},
"nvfp4_quantize": {
"7.5": [],
Expand All @@ -521,6 +542,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cuda", "cute-dsl"],
"10.3": ["cuda", "cute-dsl"],
"12.0": ["cuda", "cute-dsl"],
"12.1": ["cuda", "cute-dsl"],
},
"nvfp4_batched_quantize": {
"7.5": [],
Expand All @@ -531,6 +553,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cuda"],
"10.3": ["cuda"],
"12.0": ["cuda"],
"12.1": ["cuda"],
},
# SAMPLING
"softmax": {
Expand All @@ -542,6 +565,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cuda"],
"10.3": ["cuda"],
"12.0": ["cuda"],
"12.1": ["cuda"],
},
"sampling_from_probs": {
"7.5": ["cuda"],
Expand All @@ -552,6 +576,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cuda"],
"10.3": ["cuda"],
"12.0": ["cuda"],
"12.1": ["cuda"],
},
"sampling_from_logits": {
"7.5": ["cuda"],
Expand All @@ -562,6 +587,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cuda"],
"10.3": ["cuda"],
"12.0": ["cuda"],
"12.1": ["cuda"],
},
"top_k_sampling_from_probs": {
"7.5": ["cuda"],
Expand All @@ -572,6 +598,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cuda"],
"10.3": ["cuda"],
"12.0": ["cuda"],
"12.1": ["cuda"],
},
"top_p_sampling_from_probs": {
"7.5": ["cuda"],
Expand All @@ -582,6 +609,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cuda"],
"10.3": ["cuda"],
"12.0": ["cuda"],
"12.1": ["cuda"],
},
"top_k_top_p_sampling_from_probs": {
"7.5": ["cuda"],
Expand All @@ -592,6 +620,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cuda"],
"10.3": ["cuda"],
"12.0": ["cuda"],
"12.1": ["cuda"],
},
"top_k_top_p_sampling_from_logits": {
"7.5": ["cuda"],
Expand All @@ -602,6 +631,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cuda"],
"10.3": ["cuda"],
"12.0": ["cuda"],
"12.1": ["cuda"],
},
"min_p_sampling_from_probs": {
"7.5": ["cuda"],
Expand All @@ -612,6 +642,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cuda"],
"10.3": ["cuda"],
"12.0": ["cuda"],
"12.1": ["cuda"],
},
"top_k_renorm_probs": {
"7.5": ["cuda"],
Expand All @@ -622,6 +653,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cuda"],
"10.3": ["cuda"],
"12.0": ["cuda"],
"12.1": ["cuda"],
},
"top_p_renorm_probs": {
"7.5": ["cuda"],
Expand All @@ -632,6 +664,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cuda"],
"10.3": ["cuda"],
"12.0": ["cuda"],
"12.1": ["cuda"],
},
"top_k_mask_logits": {
"7.5": ["cuda"],
Expand All @@ -642,6 +675,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cuda"],
"10.3": ["cuda"],
"12.0": ["cuda"],
"12.1": ["cuda"],
},
"chain_speculative_sampling": {
"7.5": ["cuda"],
Expand All @@ -652,6 +686,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cuda"],
"10.3": ["cuda"],
"12.0": ["cuda"],
"12.1": ["cuda"],
},
"top_k": {
"7.5": ["cuda"],
Expand All @@ -662,6 +697,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cuda"],
"10.3": ["cuda"],
"12.0": ["cuda"],
"12.1": ["cuda"],
},
"top_k_page_table_transform": {
"7.5": ["cuda"],
Expand All @@ -672,6 +708,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cuda"],
"10.3": ["cuda"],
"12.0": ["cuda"],
"12.1": ["cuda"],
},
"top_k_ragged_transform": {
"7.5": ["cuda"],
Expand All @@ -682,6 +719,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cuda"],
"10.3": ["cuda"],
"12.0": ["cuda"],
"12.1": ["cuda"],
},
# ROPE
"apply_rope": {
Expand All @@ -693,6 +731,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cuda"],
"10.3": ["cuda"],
"12.0": ["cuda"],
"12.1": ["cuda"],
},
"apply_rope_pos_ids": {
"7.5": ["cuda"],
Expand All @@ -703,6 +742,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cuda"],
"10.3": ["cuda"],
"12.0": ["cuda"],
"12.1": ["cuda"],
},
"apply_llama31_rope": {
"7.5": ["cuda"],
Expand All @@ -713,6 +753,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cuda"],
"10.3": ["cuda"],
"12.0": ["cuda"],
"12.1": ["cuda"],
},
"apply_llama31_rope_pos_ids": {
"7.5": ["cuda"],
Expand All @@ -723,6 +764,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cuda"],
"10.3": ["cuda"],
"12.0": ["cuda"],
"12.1": ["cuda"],
},
"apply_rope_with_cos_sin_cache": {
"7.5": ["cuda"],
Expand All @@ -733,6 +775,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cuda"],
"10.3": ["cuda"],
"12.0": ["cuda"],
"12.1": ["cuda"],
},
"mla_rope_quantize_fp8": {
"7.5": [],
Expand All @@ -743,6 +786,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cuda"],
"10.3": ["cuda"],
"12.0": ["cuda"],
"12.1": ["cuda"],
},
"rope_quantize_fp8": {
"7.5": [],
Expand All @@ -753,6 +797,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cuda"],
"10.3": ["cuda"],
"12.0": ["cuda"],
"12.1": ["cuda"],
},
"rope_quantize_fp8_append_paged_kv_cache": {
"7.5": [],
Expand All @@ -763,6 +808,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.0": ["cuda"],
"10.3": ["cuda"],
"12.0": ["cuda"],
"12.1": ["cuda"],
},
# MAMBA
"selective_state_update": {
Expand All @@ -775,6 +821,7 @@ def dtype_str_to_torch_dtype(dtype_str):
"10.3": ["flashinfer", "triton"],
"11.0": ["flashinfer", "triton"],
"12.0": ["flashinfer", "triton"],
"12.1": ["flashinfer", "triton"],
},
}

Expand Down
Loading