Restructure attention benchmarks and update benchmarking artifacts (backend 9.26.0.17, frontend b21cd3f)

README.md

@@ -57,13 +57,13 @@ Contributor credits for these OSS CuTe DSL kernels are listed in [Acknowledgemen
 
 #### Llama 3.1 style Forward and Bprop with causal masking (GB300)
 <p align="center">
-  <img src="https://github.com/NVIDIA/cudnn-frontend/blob/main/benchmark/sdpa_benchmark_training/results/llama3.1/gb300/llama3.1_top_left.png" alt="Llama 3.1 SDPA Benchmark on GB300 (only cuDNN)" width="600"/>
+  <img src="https://github.com/NVIDIA/cudnn-frontend/blob/main/benchmark/attention_training/results/llama3.1/gb300/llama3.1_top_left.png" alt="Llama 3.1 SDPA Benchmark on GB300 (only cuDNN)" width="600"/>
 </p>
 
 #### Deepseek v3 style Forward and Bprop with causal masking (GB300)
 
 <p align="center">
-  <img src="https://github.com/NVIDIA/cudnn-frontend/blob/main/benchmark/sdpa_benchmark_training/results/dsv3/gb300/dsv3_top_left.png" alt="DSv3 SDPA Benchmark on GB300 (only cuDNN)" width="600"/>
+  <img src="https://github.com/NVIDIA/cudnn-frontend/blob/main/benchmark/attention_training/results/dsv3/gb300/dsv3_top_left.png" alt="DSv3 SDPA Benchmark on GB300 (only cuDNN)" width="600"/>
 </p>
 
 ## Key Features

benchmark/attention_inference/README.md

@@ -1,7 +1,7 @@
 # Attention Inference Benchmark
 
 Benchmarks attention for **inference**, split into two phases (mirroring
-forward/backward in `../sdpa_benchmark_training`):
+forward/backward in `../attention_training`):
 
 - **context** — prefill, in two kinds, both reported in **TFLOPS**:
   - *full*: `s_q == s_kv`, contiguous Q/K/V, compute-bound;
@@ -52,7 +52,7 @@ into the fp8 graph — those cases record as unsupported).
 MLA models run **absorbed** in generation (`kind="mla_absorbed"`: K reads the
 full record, V a leading slice of the *same* record, so KV bytes are counted
 once) and **unabsorbed** in prefill — which is dense training-style attention
-and lives in `../sdpa_benchmark_training`.
+and lives in `../attention_training`.
 
 ## Usage
 

benchmark/attention_inference/configs/auto_regressive_dit.py

@@ -10,7 +10,7 @@
 is bidirectional: context_causal=False). Full prefill covers the self-
 attention over the whole clip, and generation sweeps the standard MTP widths
 against the cached clip for completeness. The training suite carries its own
-forward-only sweep of the same model (`sdpa_benchmark_training`).
+forward-only sweep of the same model (`attention_training`).
 """
 
 from ..config_types import InferenceBenchmarkConfig, ModelPreset

benchmark/attention_inference/configs/qwen3vl_vit.py

@@ -0,0 +1,55 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""
+Qwen3-VL Vision Encoder (ViT) self-attention — inference.
+
+The vision tower's self-attention is bidirectional over the patchified image
+tokens and runs forward-only: pure context-phase (full prefill) work, so this
+config sweeps ONLY the context phase — an encoder has no KV cache to decode
+against, so there is no generation phase and no kv-cache dtype axis.
+
+Architecture (vision tower): 16 MHA heads, head_dim 72 zero-padded to 80
+(fp8 kernels require 16-byte-aligned head dims; production integrations run
+the padded contract, and reported TFLOPS count d=80). Sequence lengths are
+per-image patch-grid token counts from a production inference trace
+(94x94 .. 250x250 grids), spanning that trace's per-forward FLOPs
+distribution from the 10th to the 99th percentile. The tower is not
+head-shardable in deployments (whole-model per device), so no TP sweep.
+
+Usage:
+    python -m benchmark.attention_inference.runner --config qwen3vl_vit
+"""
+
+from ..config_types import InferenceBenchmarkConfig, ModelPreset
+
+QWEN3VL_VIT = ModelPreset(
+    name="qwen3vl_vit",
+    num_q_heads=16,
+    num_kv_heads=16,
+    head_dim=80,
+)
+
+CONFIG = InferenceBenchmarkConfig(
+    name="qwen3vl_vit",
+    models=[QWEN3VL_VIT],
+    context_seqlens=[
+        8836,   # 94x94 patch grid
+        15376,  # 124x124 (most frequent single-image forward)
+        24336,  # 156x156
+        35344,  # 188x188 (FLOPs-median forward)
+        47376,  # non-square grid (e.g. 168x282)
+        62500,  # 250x250
+    ],
+    context_chunked_shapes=[],   # encoder: no chunked prefill against a cache
+    generation_shapes=[],        # encoder: no decode phase
+    context_causal=False,        # bidirectional ViT self-attention
+    # bf16 only: this suite expresses fp8 solely as the generation-phase
+    # kv-cache axis (the fp8 attention graph), and an encoder has no
+    # generation phase — the context path has no fp8 route today. The training
+    # suite's fp8 forward numbers for this model are dropped by the move;
+    # restoring them here needs fp8-context support in
+    # benchmark_single_attention first.
+    data_types=["bfloat16"],
+    backends=["cudnn", "cudnn_oss"],
+)