5080deploy: rebuild on upstream main 9c635a15 (torch 2.14 + NVFP4 QSA spill fix, -20% cold TTFT)

benchmarks/README.md

@@ -27,5 +27,67 @@ batch size x miss rate.
 python benchmarks/bench_offload_cache_copy.py
 ```
 
+**`bench_scored_ensure.py`** — synthetic (no checkpoint): us per decode `ensure` call at a
+fixed miss count, flashlib `lru_ensure` vs the `--moe-cache-policy` kernels (lru, kd, kdfb,
+rule), from a CUDA graph of two decode steps on a full cache. `--stats` times it with
+`--moe-collect-stats` accumulation on.
+
+```bash
+python benchmarks/bench_scored_ensure.py --slots 1650 --k 10,20
+```
+
+**`bench_moe_copy_overlap.py`** — synthetic (no checkpoint): `FREETOKEN_MOE_COPY_OVERLAP`
+off vs on for Qwen4ExpMoE decode layers at the Qwen3.8-Flash-Next expert geometry, as CUDA
+graphs. Checks the outputs are bitwise identical, times replays ABAB, and profiles the copy's
+start after the ensure and how much of the side-stream shared expert runs inside it.
+
+```bash
+python benchmarks/bench_moe_copy_overlap.py --bs 1 --slots 160
+```
+
+**`bench_moe_prefetch.py`** — synthetic (no checkpoint): `FREETOKEN_MOE_PREFETCH=measure`
+cost. The router-lookahead predictor alone per layer (GEMV `[bs, 2560] x [2560, 512]`, top-K
+select, and the count kernel), then prefetch off vs measure on a real-geometry Qwen4ExpMoE decode
+stack (512 experts, rule eviction, copy overlap on) with forced routing that misses exactly m
+experts per layer, as CUDA graphs. `--on` times whole 48-layer decode steps with prefetch off vs
+on instead: a GEMV per layer stands in for attention (`--filler-us`), each layer prefetches 3 or
+4 candidates of which `--useful` are its misses, and `--delay-us` holds every prefetch copy back
+to price a late one. It reports ms/step, demand misses/step and the late fraction. `--steady
+--trace DIR` runs steady-state decode on a recorded routing trace instead (experts.npy,
+logit_idx.npy, logit_val.npy, req_index.npy): no cache restore, so wrong prefetches and evictions
+carry over, and a noisy copy of each layer's traced logits (`--sigma`) stands in for the
+lookahead. Arms such as `off,measure,on,on:4/4` (budgets before GDN / QSA layers) replay step by
+step in turn, each on its own copy of the cache state; it adds pollution (non-resident demands
+per layer against off's misses) and the late fraction per layer kind. Slots and sigma were
+calibrated to production's measure line (misses/layer ~4.3, precision ~0.56) at 24 layers.
+
+```bash
+python benchmarks/bench_moe_prefetch.py --bs 1,2 --misses 0,4
+python benchmarks/bench_moe_prefetch.py --on --bs 1,2 --misses 4 --delay-us 0,50,150,300
+python benchmarks/bench_moe_prefetch.py --steady --trace DIR --bs 1,2 --steady-slots 740 --sigma 0.55 \
+    --arms off,measure,on,on:empty,on:3/3,on:4/4,on:4/5
+```
+
+**`bench_expert_copy_corun.py`** — synthetic (no checkpoint): the fused expert copy kernel
+alone and beside compute, on registered host banks with the qwen4_exp NVFP4 row sizes. Sweeps
+`fast_index_copy_multi` (threads x blocks/bank) and `fast_index_copy_multi_slim`
+(blocks x threads x unroll) over 1..10-expert plans (GB/s, "hot" back-to-back and "cold"
+spacer-separated copies), then reports the stretch sigma of bf16/fp8 GEMVs, a small matmul and
+the NVFP4 routed decode GEMV while a copy loop runs on another stream. `--json` keeps the raw
+per-trial times.
+
+```bash
+python benchmarks/bench_expert_copy_corun.py --trials 5 --json copy_corun.json
+```
+
 For host RAM vs PCIe bandwidth and the offload/hybrid backend pick, use `ft bench bw`
 instead — it writes the JSON profile the engine reads.
+
+**`bench_kv_quant.py`** compares BF16, FP8 and NVFP4 KV storage bytes, one-step
+scatter latency and paged decode latency on synthetic inputs. No checkpoint is
+required. Keep the GPU idle and use identical arguments for A/B comparisons;
+this does not measure model quality or end-to-end serving throughput.
+
+```bash
+PYTHONPATH=python:. uv run python benchmarks/bench_kv_quant.py --lengths 1024,8192,32768
+```