Add per-layer OI/CF roofline validation artifacts

.gitattributes

@@ -59,3 +59,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+per_layer_oi_cf/ncu/Llama-3.1-8B_prefill_bs1_8layers.ncu-rep filter=lfs diff=lfs merge=lfs -text
+per_layer_oi_cf/ncu/Llama-3.1-8B_prefill_bs1_v2.ncu-rep filter=lfs diff=lfs merge=lfs -text

per_layer_oi_cf/README.md

@@ -0,0 +1,42 @@
+# Per-Layer OI/CF Roofline — Llama-3.1-8B on H100 (Section 4.4)
+
+Per-component operational intensity (OI) and capacity footprint (CF) roofline
+decomposition for a Llama-3.1-8B decoder layer on a single H100-SXM5-80GB
+(bf16, TP=1). Backs Section 4.4 "Layer-Level Roofline Decomposition" and
+Figure `per_layer_oi_cf_8b_h100.pdf`.
+
+## Files
+
+- `analytical/` — Per-component OI computed from model dimensions.
+  B=1 prefill (C=1, memory-limited) and B=80 prefill (C=80, compute-saturated).
+- `ncu/` — NCU v2024.3.2 kernel traces (`--set full`) on 8 decoder layers
+  (indices 1–8) at bs=1 prefill. Clean run, 52 us total GPU time across
+  8 unique kernel types. Reproduce with:
+  ```
+  CUDA_VISIBLE_DEVICES=0 ncu --set full --csv --profile-from-start no \
+    --target-processes all -o output \
+    python3 scripts/roofline/_ncu_target_simple.py
+  ```
+- `per_layer_oi_cf_8b_h100.pdf` — Two-panel figure: (a) per-component OI on
+  H100 roofline, (b) capacity footprint sweep C=1..80.
+- `per_layer_table.tex` — LaTeX table classifying each component as
+  compute-bound (OI > 295 FLOP/byte) or bandwidth-bound.
+
+## Key findings
+
+| Component | Wt (MB) | OI (C=1) | OI (C=80) | Bound |
+|-----------|---------|----------|-----------|-------|
+| GEMM projections (Q,K,V,O,Gate,Up,Down) | 436 | 315–441 | 803–2956 | COMPUTE |
+| RMSNorm (in/out), RoPE, SiLU, Gate×Up | — | 0–1 | 0–1 | BANDWIDTH |
+| Flash attention | — | 64 | 64 | BANDWIDTH |
+| **Layer total** | **436** | **330** | **878** | **COMPUTE** |
+
+NCU validation: at C=1, GEMM = 10.8% of GPU time, memory ops = 65%,
+confirming the layer is barely compute-bound (OI=330 vs ridge=295).
+
+System capacity-bound at 32K context: C ≥ 16 (KV cache + weights > 80 GB HBM).
+
+## Citation
+
+AgentPerfBench: A Benchmarking and Evaluation Suite for Inference Performance
+of Agentic LLMs. NeurIPS 2026.

per_layer_oi_cf/analytical/llama8b_analytical_prefill_B1.json

@@ -0,0 +1,110 @@
+{
+  "B": 1,
+  "S": 512,
+  "phase": "prefill",
+  "d": 4096,
+  "h": 32,
+  "kv_h": 8,
+  "intermediate": 14336,
+  "head_dim": 128,
+  "q_len": 512,
+  "kv_len": 512,
+  "total_flops": 227707715584,
+  "total_bytes": 689963008,
+  "oi": 330.0,
+  "ridge_point": 295.0,
+  "overall_bound": "compute",
+  "components": [
+    {
+      "name": "rmsnorm_in",
+      "flops": 10485760,
+      "bytes": 8388608,
+      "oi": 1.2,
+      "bound": "memory"
+    },
+    {
+      "name": "q_proj",
+      "flops": 17179869184,
+      "bytes": 41943040,
+      "oi": 409.6,
+      "bound": "compute"
+    },
+    {
+      "name": "k_proj",
+      "flops": 4294967296,
+      "bytes": 13631488,
+      "oi": 315.1,
+      "bound": "compute"
+    },
+    {
+      "name": "v_proj",
+      "flops": 4294967296,
+      "bytes": 13631488,
+      "oi": 315.1,
+      "bound": "compute"
+    },
+    {
+      "name": "rope",
+      "flops": 16777216,
+      "bytes": 12582912,
+      "oi": 1.3,
+      "bound": "memory"
+    },
+    {
+      "name": "attention",
+      "flops": 4294967296,
+      "bytes": 67108864,
+      "oi": 64.0,
+      "bound": "memory"
+    },
+    {
+      "name": "o_proj",
+      "flops": 17179869184,
+      "bytes": 41943040,
+      "oi": 409.6,
+      "bound": "compute"
+    },
+    {
+      "name": "rmsnorm_post",
+      "flops": 10485760,
+      "bytes": 8388608,
+      "oi": 1.2,
+      "bound": "memory"
+    },
+    {
+      "name": "gate_proj",
+      "flops": 60129542144,
+      "bytes": 136314880,
+      "oi": 441.1,
+      "bound": "compute"
+    },
+    {
+      "name": "up_proj",
+      "flops": 60129542144,
+      "bytes": 136314880,
+      "oi": 441.1,
+      "bound": "compute"
+    },
+    {
+      "name": "silu",
+      "flops": 29360128,
+      "bytes": 29360128,
+      "oi": 1.0,
+      "bound": "memory"
+    },
+    {
+      "name": "gate_mul",
+      "flops": 7340032,
+      "bytes": 44040192,
+      "oi": 0.2,
+      "bound": "memory"
+    },
+    {
+      "name": "down_proj",
+      "flops": 60129542144,
+      "bytes": 136314880,
+      "oi": 441.1,
+      "bound": "compute"
+    }
+  ]
+}

per_layer_oi_cf/analytical/llama8b_analytical_prefill_B80.json

@@ -0,0 +1,110 @@
+{
+  "B": 80,
+  "S": 512,
+  "phase": "prefill",
+  "d": 4096,
+  "h": 32,
+  "kv_h": 8,
+  "intermediate": 14336,
+  "head_dim": 128,
+  "q_len": 512,
+  "kv_len": 512,
+  "total_flops": 18216617246720,
+  "total_bytes": 20736638976,
+  "oi": 878.5,
+  "ridge_point": 295.0,
+  "overall_bound": "compute",
+  "components": [
+    {
+      "name": "rmsnorm_in",
+      "flops": 838860800,
+      "bytes": 671088640,
+      "oi": 1.2,
+      "bound": "memory"
+    },
+    {
+      "name": "q_proj",
+      "flops": 1374389534720,
+      "bytes": 704643072,
+      "oi": 1950.5,
+      "bound": "compute"
+    },
+    {
+      "name": "k_proj",
+      "flops": 343597383680,
+      "bytes": 427819008,
+      "oi": 803.1,
+      "bound": "compute"
+    },
+    {
+      "name": "v_proj",
+      "flops": 343597383680,
+      "bytes": 427819008,
+      "oi": 803.1,
+      "bound": "compute"
+    },
+    {
+      "name": "rope",
+      "flops": 1342177280,
+      "bytes": 1006632960,
+      "oi": 1.3,
+      "bound": "memory"
+    },
+    {
+      "name": "attention",
+      "flops": 343597383680,
+      "bytes": 5368709120,
+      "oi": 64.0,
+      "bound": "memory"
+    },
+    {
+      "name": "o_proj",
+      "flops": 1374389534720,
+      "bytes": 704643072,
+      "oi": 1950.5,
+      "bound": "compute"
+    },
+    {
+      "name": "rmsnorm_post",
+      "flops": 838860800,
+      "bytes": 671088640,
+      "oi": 1.2,
+      "bound": "memory"
+    },
+    {
+      "name": "gate_proj",
+      "flops": 4810363371520,
+      "bytes": 1627389952,
+      "oi": 2955.9,
+      "bound": "compute"
+    },
+    {
+      "name": "up_proj",
+      "flops": 4810363371520,
+      "bytes": 1627389952,
+      "oi": 2955.9,
+      "bound": "compute"
+    },
+    {
+      "name": "silu",
+      "flops": 2348810240,
+      "bytes": 2348810240,
+      "oi": 1.0,
+      "bound": "memory"
+    },
+    {
+      "name": "gate_mul",
+      "flops": 587202560,
+      "bytes": 3523215360,
+      "oi": 0.2,
+      "bound": "memory"
+    },
+    {
+      "name": "down_proj",
+      "flops": 4810363371520,
+      "bytes": 1627389952,
+      "oi": 2955.9,
+      "bound": "compute"
+    }
+  ]
+}

per_layer_oi_cf/ncu/Llama-3.1-8B_prefill_bs1_8layers.ncu-rep

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6feef6e33821aee9fa5f0a3cc174cba6947106644f3ebe77f08e58fc98e69e1e
+size 72020417

per_layer_oi_cf/ncu/Llama-3.1-8B_prefill_bs1_v2.ncu-rep

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a248061d85e6ce523e51d258fc50fad7140e72eeb6cd5dce468839a289c3517e
+size 110769455

per_layer_oi_cf/per_layer_table.tex

@@ -0,0 +1,29 @@
+\begin{table}[t]
+\centering
+\caption{Per-component operational intensity for a Llama-3.1-8B decoder layer on H100 (bf16, TP=1). Ridge = 295 FLOP/byte. C = compute-bound, BW = memory-bandwidth-bound. Capacity-bound arises at 8K+ contexts when KV cache exceeds HBM.}
+\label{tab:per-component-oi}
+\small
+\begin{tabular}{lrrrrr}
+\toprule
+Component & Wt (MB) & OI (C=1) & OI (C=80) & Bound (C=1) & Bound (C=80) \\
+\midrule
+RMS & \textemdash & 1 & 1 & BW & BW \\
+Q & 34 & 410 & 1950 & C & C \\
+K & 8 & 315 & 803 & C & C \\
+V & 8 & 315 & 803 & C & C \\
+RoPE & \textemdash & 1 & 1 & BW & BW \\
+Attn & \textemdash & 64 & 64 & BW & BW \\
+O & 34 & 410 & 1950 & C & C \\
+RMS & \textemdash & 1 & 1 & BW & BW \\
+Gate & 117 & 441 & 2956 & C & C \\
+Up & 117 & 441 & 2956 & C & C \\
+SiLU & \textemdash & 1 & 1 & BW & BW \\
+Gate×Up & \textemdash & 0 & 0 & BW & BW \\
+Down & 117 & 441 & 2956 & C & C \\
+\midrule
+\textbf{Layer total} & 436 & 330 & 878 & C & C \\
+\bottomrule
+\end{tabular}
+\vspace{3pt}
+{\footnotesize KV capacity thresholds: 8K tok $\rightarrow$ C$\ge$57; 32K tok $\rightarrow$ C$\ge$14; 128K tok $\rightarrow$ C$\ge$3.}
+\end{table}