initial upload: turboquant KV cache benchmarks on RTX 4060 Ti 8GB

README.md

@@ -0,0 +1,108 @@
+---
+dataset_info:
+  features:
+    - name: config
+      dtype: string
+    - name: runtime
+      dtype: string
+    - name: kv_cache_k
+      dtype: string
+    - name: kv_cache_v
+      dtype: string
+    - name: context_k
+      dtype: int64
+    - name: np_slots
+      dtype: int64
+    - name: graph_splits
+      dtype: int64
+    - name: kv_vram_mib
+      dtype: int64
+    - name: checkpoint_mode
+      dtype: string
+    - name: agent
+      dtype: string
+    - name: system_prompt_tokens
+      dtype: int64
+    - name: context_depth_k
+      dtype: float64
+    - name: decode_tok_s
+      dtype: float64
+    - name: prompt_eval_tok_s
+      dtype: float64
+    - name: test_type
+      dtype: string
+    - name: notes
+      dtype: string
+license: mit
+tags:
+  - benchmark
+  - local-llm
+  - 8gb-vram
+  - turboquant
+  - moe-offload
+  - kv-cache
+  - rtx-4060-ti
+  - qwen
+  - practitioner-tested
+size_categories:
+  - n<1K
+---
+
+# RTX 4060 Ti 8GB — turboquant KV cache benchmark (Qwen3.6-35B-A3B)
+
+practitioner-tested benchmarks of turboquant KV cache types vs standard llama.cpp on an RTX 4060 Ti 8GB with 32GB DDR5-6000 RAM.
+
+## hardware
+
+| component | spec |
+|-----------|------|
+| GPU | NVIDIA RTX 4060 Ti, 8 GB VRAM |
+| CPU | AMD Ryzen 5 7600X (6c/12t) |
+| RAM | 32 GB DDR5-6000 dual-channel |
+| OS | Windows 11 + WSL2 Ubuntu 26.04 |
+
+## model
+
+Qwen3.6-35B-A3B-UD-Q4_K_M (22.1 GB). hybrid SSM+attention architecture — 10/40 layers use attention, 30/40 use Gated Delta Net (recurrent). MoE with partial offload: -ncmoe 30 (10 layers experts on GPU, 30 on CPU).
+
+## what this dataset covers
+
+1. **turboquant KV cache types** — turbo2, turbo3, turbo4 vs standard q4_0 and q8_0. turboquant is a llama.cpp fork (github.com/TheTom/llama-cpp-turboquant) that adds aggressive KV cache compression with auto-asymmetric behaviour (K stays q8_0 for high GQA ratios, only V is compressed).
+
+2. **graph splits** — the hidden speed determinant on 8GB VRAM. standard llama.cpp V q4_0 at 64K context produces 82 graph splits (GPU-CPU transfers per step) vs 62 at 32K. turboquant turbo2/turbo3 keep 62 splits at 64K. the split count matters more than raw VRAM usage at this boundary.
+
+3. **context checkpoint impact** — llama.cpp creates context checkpoints every 8192 tokens during prefill (~63 MiB each in VRAM). at 26K context, 5 checkpoints = 315 MiB, pushing past the 7 GB VRAM cliff. disabling with --checkpoint-every-n-tokens -1 eliminates the cliff entirely. this is a novel finding — I have not seen it documented elsewhere.
+
+4. **agent harness overhead** — Hermes Agent v0.13.0 (~13.5K token system prompt) vs Pi coding agent v0.74.0 (~1.9K tokens). same decode speed, 7x smaller system prompt = 13K more usable context.
+
+5. **speed curve at depth** — decode speed from 3.5K to 50K context with checkpoints disabled. smooth degradation (~2 tok/s per 10K tokens) from attention scaling on the 10 attention layers. no cliff.
+
+## key findings
+
+- **turbo3 is the sweet spot for 8GB VRAM.** same speed as turbo2 (35 tok/s) with better V cache compression. turbo4 is too aggressive (13.93 tok/s — decompression overhead kills it).
+- **graph splits > raw VRAM** as the speed determinant at this boundary. 62 splits = fast, 82 splits = broken.
+- **context checkpoints are a hidden VRAM eater.** disabling them is free on Qwen3.6 (hybrid architecture invalidates them anyway).
+- **final config:** 64K context, 28-34 tok/s across all depths, no cliff.
+
+## recommended config
+
+```
+-ngl 999 -ncmoe 30 -c 65536 -np 1 -fa on --cache-type-k q8_0 --cache-type-v turbo3 --no-cache-prompt --checkpoint-every-n-tokens -1
+```
+
+## methodology
+
+- all tests on same hardware, same Q4_K_M quant, same -ncmoe 30 partial offload
+- curl tests: single 300-token completion request (warm run)
+- agent tests: real multi-step tasks via Hermes or Pi
+- decode speed from llama-server slot print_timing logs
+- tested 2026-05-11 and 2026-05-12
+
+## related datasets
+
+- [witcheer/windows-rtx-4060ti-8gb-bench-2026-05](https://huggingface.co/datasets/witcheer/windows-rtx-4060ti-8gb-bench-2026-05) — dense model benchmarks
+- [witcheer/windows-rtx-4060ti-8gb-moe-offload-bench-2026-05](https://huggingface.co/datasets/witcheer/windows-rtx-4060ti-8gb-moe-offload-bench-2026-05) — MoE offload benchmarks
+
+## collection
+
+[8GB VRAM local LLMs — practitioner-tested](https://huggingface.co/collections/witcheer/8gb-vram-local-llms-practitioner-tested)

benchmark.csv

@@ -0,0 +1,19 @@
+config,runtime,kv_cache_k,kv_cache_v,context_k,np_slots,graph_splits,kv_vram_mib,checkpoint_mode,agent,system_prompt_tokens,context_depth_k,decode_tok_s,prompt_eval_tok_s,test_type,notes
+baseline_hermes,llama.cpp,q8_0,q8_0,32768,4,62,1360,default,hermes,13500,3.5,31.4,,hermes_round,original smoke test config
+phase1_np1,llama.cpp,q8_0,q8_0,32768,1,62,340,default,none,,3.5,34.94,,curl,added -np 1
+phase2_q4v_64k,llama.cpp,q8_0,q4_0,65536,1,82,520,default,none,,3.5,31.22,,curl,V cache q4_0 at 64K
+phase2_q4v_64k,llama.cpp,q8_0,q4_0,65536,1,82,520,default,hermes,13500,14.0,9.0,,hermes_round,82 splits killed Hermes speed
+phase2_q4v_48k,llama.cpp,q8_0,q4_0,49152,1,82,390,default,none,,3.5,30.52,,curl,48K still has 82 splits
+turboquant_turbo4,turboquant,q8_0_auto,turbo4,65536,1,62,510,default,none,,3.5,13.93,,curl,auto-asymmetric upgraded K to q8_0; turbo4 decompression too aggressive
+turboquant_turbo2,turboquant,q8_0,turbo2,65536,1,62,510,default,none,,3.5,35.13,,curl,turbo2 keeps 62 splits at 64K
+turboquant_turbo2,turboquant,q8_0,turbo2,65536,1,62,510,default,hermes,13500,14.3,33.59,,hermes_round,confirmed speed holds in Hermes
+turboquant_turbo2,turboquant,q8_0,turbo2,65536,1,62,510,default,hermes,13500,14.3,35.76,,hermes_round,second call same session
+turboquant_turbo3,turboquant,q8_0,turbo3,65536,1,62,510,default,none,,3.5,35.17,,curl,turbo3 matches turbo2 speed with better compression
+turboquant_turbo3,turboquant,q8_0,turbo3,65536,1,62,510,default,pi,1900,3.5,34.18,,pi_round,Pi system prompt 7x smaller than Hermes
+turboquant_turbo3,turboquant,q8_0,turbo3,65536,1,62,510,default,pi,1900,17.2,31.17,,pi_round,speed holds at 17K context
+turboquant_turbo3,turboquant,q8_0,turbo3,65536,1,62,510,default,pi,1900,26.5,14.64,88,pi_round,VRAM cliff from checkpoint accumulation
+turboquant_turbo3,turboquant,q8_0,turbo3,65536,1,62,510,disabled,none,,3.5,34.25,370,curl,checkpoints disabled via --checkpoint-every-n-tokens -1
+turboquant_turbo3,turboquant,q8_0,turbo3,65536,1,62,510,disabled,none,,20.0,31.54,461,curl,no cliff at 20K
+turboquant_turbo3,turboquant,q8_0,turbo3,65536,1,62,510,disabled,none,,26.0,30.26,443,curl,no cliff at 26K — was 14.64 with checkpoints
+turboquant_turbo3,turboquant,q8_0,turbo3,65536,1,62,510,disabled,none,,35.0,29.62,455,curl,smooth degradation continues
+turboquant_turbo3,turboquant,q8_0,turbo3,65536,1,62,510,disabled,none,,50.0,27.71,450,curl,50K context at 28 tok/s — no cliff