feat: add CPU benchmark tool with context length sweep

Update README.md

Author: huangzhenhua111

.gitignore

@@ -40,3 +40,12 @@ autotuner.log
 /models/
 # Keep source model adapters tracked
 !tools/mllm-llm-benchmark/models/
+
+# ---- bench artifacts (generated / local) ----
+perf.data
+perf.data.old
+bench_context/
+snapshots/
+*.bak
+*.bak.*
+bench_mem_context.csv

bench_artifacts/README.md

@@ -0,0 +1,39 @@
+# [Bench] CPU Context Sweep Benchmark
+
+CPU-only benchmark for measuring prefill/decode latency and memory 
+across context lengths (256–4096) on x86_64.
+
+`sweep_context_v2.sh` runs the benchmark at each context length in two 
+modes (decode_heavy, prefill_ttft) and records: TTFT, prefill latency, 
+decode per-token latency, peak RSS, KV cache estimate, and GEMM shapes 
+(via `MLLM_MATMUL_SHAPE_LOG=1`).
+
+- `data/` — raw CSVs
+- `plots/` — generated figures
+
+## Quick Repro (x86_64 example)
+
+```bash
+# 1. Build 
+cmake -S . -B build -G Ninja \
+  -DCMAKE_BUILD_TYPE=RelWithDebInfo \
+  -DBUILD_TESTING=OFF \
+  -DMLLM_ENABLE_TOOLS=ON \
+  -DCMAKE_CXX_FLAGS="-march=native"
+
+cmake --build build -j"$(nproc)"
+
+# 2. Run Context Sweep & View Shape Log
+chmod +x sweep_context_v2.sh
+export BIN=./build/bin/mllm-llm-benchmark
+export MLLM_MATMUL_SHAPE_LOG=1
+
+# Run a quick test
+./sweep_context_v2.sh
+```
+
+Outputs:
+
+* `bench_context/context_sweep_v2.csv`
+* <img width="2879" height="1366" alt="image" src="https://github.com/user-attachments/assets/46a4cb2b-6a22-4fba-81e2-dccf41bf4357" />
+* <img width="2837" height="1437" alt="image" src="https://github.com/user-attachments/assets/3468b710-c7c8-4bab-bfe6-afaf6826d5ed" />

bench_artifacts/data/context_sweep_v2.csv

@@ -0,0 +1,10 @@
+ts,git,arch,model,mode,cl,pp,tg,threads,ttft_ms,prefill_ms,decode_ms,decode_ms_per_tok,peak_rss_kb,kv_est_kb
+2026-02-02T17:28:18+08:00,b17c8f55eea8,x86_64,tiny_llama,prefill_ttft,256,254,2,8,8476.504,8476.504,169.632,84.816,4207436,5632
+2026-02-02T17:28:18+08:00,b17c8f55eea8,x86_64,tiny_llama,decode_heavy,512,256,256,8,7719.204,7719.2046,173.43939,0.6774976171875,4281236,11264
+2026-02-02T17:28:18+08:00,b17c8f55eea8,x86_64,tiny_llama,prefill_ttft,512,510,2,8,18217.402,18217.402,179.518,89.759,4364668,11264
+2026-02-02T17:28:18+08:00,b17c8f55eea8,x86_64,tiny_llama,decode_heavy,1024,768,256,8,28733.297,28733.295,172.73177,0.6747334765625,4756888,22528
+2026-02-02T17:28:18+08:00,b17c8f55eea8,x86_64,tiny_llama,prefill_ttft,1024,1022,2,8,50484.66,50484.66,199.94199,99.970995,4841960,22528
+2026-02-02T17:28:18+08:00,b17c8f55eea8,x86_64,tiny_llama,decode_heavy,2048,1792,256,8,135609.38,135609.38,187.38985,0.7319916015625,5932308,45056
+2026-02-02T17:28:18+08:00,b17c8f55eea8,x86_64,tiny_llama,prefill_ttft,2048,2046,2,8,149802.72,149802.72,259.102,129.551,6221464,45056
+2026-02-02T17:28:18+08:00,b17c8f55eea8,x86_64,tiny_llama,decode_heavy,4096,3840,256,8,462259.72,462259.75,224.40271,0.8765730859375,7343064,90112
+2026-02-02T17:28:18+08:00,b17c8f55eea8,x86_64,tiny_llama,prefill_ttft,4096,4094,2,8,523101.5,523101.56,1793.434,896.717,7363988,90112

bench_artifacts/data/context_sweep_v2.summary.csv

@@ -0,0 +1,10 @@
+ts,git,arch,model,mode,cl,pp,tg,threads,ttft_ms,prefill_ms,decode_ms,decode_ms_per_tok,peak_rss_kb,kv_est_kb,peak_rss_gb,kv_est_mb
+2026-02-02T17:28:18+08:00,b17c8f55eea8,x86_64,tiny_llama,decode_heavy,512.0,256.0,256.0,8.0,7719.204,7719.2046,173.43939,0.6774976171875,4281236.0,11264.0,4.082904815673828,11.0
+2026-02-02T17:28:18+08:00,b17c8f55eea8,x86_64,tiny_llama,decode_heavy,1024.0,768.0,256.0,8.0,28733.297,28733.295,172.73177,0.6747334765625,4756888.0,22528.0,4.536521911621094,22.0
+2026-02-02T17:28:18+08:00,b17c8f55eea8,x86_64,tiny_llama,decode_heavy,2048.0,1792.0,256.0,8.0,135609.38,135609.38,187.38985,0.7319916015625,5932308.0,45056.0,5.657489776611328,44.0
+2026-02-02T17:28:18+08:00,b17c8f55eea8,x86_64,tiny_llama,decode_heavy,4096.0,3840.0,256.0,8.0,462259.72,462259.75,224.40271,0.8765730859375,7343064.0,90112.0,7.002891540527344,88.0
+2026-02-02T17:28:18+08:00,b17c8f55eea8,x86_64,tiny_llama,prefill_ttft,256.0,254.0,2.0,8.0,8476.504,8476.504,169.632,84.816,4207436.0,5632.0,4.012523651123047,5.5
+2026-02-02T17:28:18+08:00,b17c8f55eea8,x86_64,tiny_llama,prefill_ttft,512.0,510.0,2.0,8.0,18217.402,18217.402,179.518,89.759,4364668.0,11264.0,4.162471771240234,11.0
+2026-02-02T17:28:18+08:00,b17c8f55eea8,x86_64,tiny_llama,prefill_ttft,1024.0,1022.0,2.0,8.0,50484.66,50484.66,199.94199,99.970995,4841960.0,22528.0,4.617652893066406,22.0
+2026-02-02T17:28:18+08:00,b17c8f55eea8,x86_64,tiny_llama,prefill_ttft,2048.0,2046.0,2.0,8.0,149802.72,149802.72,259.102,129.551,6221464.0,45056.0,5.933250427246094,44.0
+2026-02-02T17:28:18+08:00,b17c8f55eea8,x86_64,tiny_llama,prefill_ttft,4096.0,4094.0,2.0,8.0,523101.5,523101.56,1793.434,896.717,7363988.0,90112.0,7.022846221923828,88.0