Dataset Viewer
Auto-converted to Parquet Duplicate
model_id
stringclasses
1 value
model_display_name
stringclasses
1 value
publisher
stringclasses
1 value
quantization
stringclasses
3 values
format
stringclasses
1 value
regime
stringclasses
2 values
hardware
dict
runtime
dict
benchmark
dict
evaluation
dict
session
dict
Qwen/Qwen3.6-35B-A3B
qwen3.6-35b-a3b
qwen
Q4_K_S
GGUF
moe_cpu_expert_offload
{ "gpu": "NVIDIA GeForce RTX 4060 Ti", "vram_gb": 8, "system_ram_gb": 32, "ram_spec": "DDR5-6000 CL36", "cpu": "AMD Ryzen 5 7600X", "platform": "Windows 11" }
{ "engine": "LM Studio", "context_length": 16384, "gpu_offload": "max", "expert_offload_to_cpu": true, "kv_cache_quant_k": "q8_0", "kv_cache_quant_v": "q8_0", "flash_attention": true, "thinking_disabled": true, "temperature": 0.7, "top_p": 0.9 }
{ "tok_per_sec": 7.4, "ttft_seconds": 1.75, "tokens_generated": 761, "vram_used_gb": null, "ram_used_gb_peak": 31, "prompt_id": "moe-explainer-500w" }
{ "quality_5pt": 4, "hit_length_target": true, "notes": "RAM-ceilinged on 32GB system. Reproduces 5.5x slower than the public X-post claim of 41 tok/sec — same author's thread requires 64GB RAM as the unstated prerequisite. GGUF repackaged by bartowski." }
{ "date": "2026-05-06T00:00:00", "run_type": "warm" }
Qwen/Qwen3.6-35B-A3B
qwen3.6-35b-a3b
qwen
Q4_K_S
GGUF
moe_cpu_expert_offload
{ "gpu": "NVIDIA GeForce RTX 4060 Ti", "vram_gb": 8, "system_ram_gb": 32, "ram_spec": "DDR5-6000 CL36", "cpu": "AMD Ryzen 5 7600X", "platform": "Windows 11" }
{ "engine": "LM Studio", "context_length": 32768, "gpu_offload": "max", "expert_offload_to_cpu": true, "kv_cache_quant_k": "q8_0", "kv_cache_quant_v": "q8_0", "flash_attention": true, "thinking_disabled": true, "temperature": 0.7, "top_p": 0.9 }
{ "tok_per_sec": 7.43, "ttft_seconds": 1.85, "tokens_generated": 796, "vram_used_gb": null, "ram_used_gb_peak": 31, "prompt_id": "moe-explainer-500w" }
{ "quality_5pt": 4, "hit_length_target": true, "notes": "Doubling context from 16K to 32K barely moved tok/sec (7.40 → 7.43). Flat curve = constant overhead dominates, not KV-attention-scan cost. Consistent with disk-paging bottleneck rather than RAM-bandwidth bottleneck. GGUF repackaged by bartowski." }
{ "date": "2026-05-06T00:00:00", "run_type": "warm" }
Qwen/Qwen3.6-35B-A3B
qwen3.6-35b-a3b
qwen
IQ4_XS
GGUF
moe_cpu_expert_offload
{ "gpu": "NVIDIA GeForce RTX 4060 Ti", "vram_gb": 8, "system_ram_gb": 32, "ram_spec": "DDR5-6000 CL36", "cpu": "AMD Ryzen 5 7600X", "platform": "Windows 11" }
{ "engine": "LM Studio", "context_length": 16384, "gpu_offload": "max", "expert_offload_to_cpu": true, "kv_cache_quant_k": "q8_0", "kv_cache_quant_v": "q8_0", "flash_attention": true, "thinking_disabled": true, "temperature": 0.7, "top_p": 0.9 }
{ "tok_per_sec": 7.4, "ttft_seconds": null, "tokens_generated": null, "vram_used_gb": null, "ram_used_gb_peak": 31, "prompt_id": "moe-explainer-500w" }
{ "quality_5pt": null, "hit_length_target": null, "notes": "Speed-only run, no quality eval captured. Smaller quant (~2GB less than Q4_K_S, 18.81GB vs 20.59GB) tested to isolate quant size as variable. Same tok/sec confirms quant size is not the lever — RAM ceiling is the binding constraint regardless of weight f...
{ "date": "2026-05-06T00:00:00", "run_type": "warm" }
Qwen/Qwen3.6-35B-A3B
qwen3.6-35b-a3b
qwen
Q4_K_M
GGUF
moe_partial_cpu_expert_offload
{ "gpu": "NVIDIA GeForce RTX 4060 Ti", "vram_gb": 8, "system_ram_gb": 32, "ram_spec": "DDR5-6000 CL36", "cpu": "AMD Ryzen 5 7600X", "platform": "WSL2 Ubuntu 26.04 on Windows 11" }
{ "engine": "llama-server (llama.cpp b9049-2496f9c14)", "context_length": 16384, "gpu_offload": "max", "expert_offload_to_cpu": true, "kv_cache_quant_k": "q8_0", "kv_cache_quant_v": "q8_0", "flash_attention": true, "thinking_disabled": false, "temperature": 0.7, "top_p": null }
{ "tok_per_sec": 29.7, "ttft_seconds": 2.44, "tokens_generated": 1024, "vram_used_gb": null, "ram_used_gb_peak": null, "prompt_id": "moe-explainer-500w" }
{ "quality_5pt": null, "hit_length_target": null, "notes": "ncmoe=32 (8 layers experts on GPU, 32 on CPU). ngl=999, threads=6. Prompt: 20.46 tok/s. Total: 36.9s. VRAM ~6649 MiB / 8187 MiB. CPU-mapped experts: 16581 MiB. Quantized by Unsloth (UD). Baseline partial offload — clean, no VRAM pressure. 4x faster than ...
{ "date": "2026-05-07T00:00:00", "run_type": "cold_server" }
Qwen/Qwen3.6-35B-A3B
qwen3.6-35b-a3b
qwen
Q4_K_M
GGUF
moe_partial_cpu_expert_offload
{ "gpu": "NVIDIA GeForce RTX 4060 Ti", "vram_gb": 8, "system_ram_gb": 32, "ram_spec": "DDR5-6000 CL36", "cpu": "AMD Ryzen 5 7600X", "platform": "WSL2 Ubuntu 26.04 on Windows 11" }
{ "engine": "llama-server (llama.cpp b9049-2496f9c14)", "context_length": 16384, "gpu_offload": "max", "expert_offload_to_cpu": true, "kv_cache_quant_k": "q8_0", "kv_cache_quant_v": "q8_0", "flash_attention": true, "thinking_disabled": false, "temperature": 0.7, "top_p": null }
{ "tok_per_sec": 32, "ttft_seconds": 1.18, "tokens_generated": 1024, "vram_used_gb": null, "ram_used_gb_peak": null, "prompt_id": "moe-explainer-500w" }
{ "quality_5pt": null, "hit_length_target": null, "notes": "ncmoe=30 (10 layers experts on GPU, 30 on CPU). ngl=999, threads=6. Prompt: 42.48 tok/s. Total: 33.2s. Sweet spot for 16K context. 7.7% faster decode than ncmoe 32. 4.3x faster than full offload." }
{ "date": "2026-05-07T00:00:00", "run_type": "cold_server" }
Qwen/Qwen3.6-35B-A3B
qwen3.6-35b-a3b
qwen
Q4_K_M
GGUF
moe_partial_cpu_expert_offload
{ "gpu": "NVIDIA GeForce RTX 4060 Ti", "vram_gb": 8, "system_ram_gb": 32, "ram_spec": "DDR5-6000 CL36", "cpu": "AMD Ryzen 5 7600X", "platform": "WSL2 Ubuntu 26.04 on Windows 11" }
{ "engine": "llama-server (llama.cpp b9049-2496f9c14)", "context_length": 16384, "gpu_offload": "max", "expert_offload_to_cpu": true, "kv_cache_quant_k": "q8_0", "kv_cache_quant_v": "q8_0", "flash_attention": true, "thinking_disabled": false, "temperature": 0.7, "top_p": null }
{ "tok_per_sec": 16.33, "ttft_seconds": 2.05, "tokens_generated": 1024, "vram_used_gb": null, "ram_used_gb_peak": null, "prompt_id": "moe-explainer-500w" }
{ "quality_5pt": null, "hit_length_target": null, "notes": "ncmoe=28 (12 layers experts on GPU, 28 on CPU). ngl=999, threads=6. Prompt: 24.38 tok/s. Total: 64.7s. VRAM pressure triggered — CUDA page faults through PCIe halved throughput. Sharp cliff, not gradual." }
{ "date": "2026-05-07T00:00:00", "run_type": "cold_server" }
Qwen/Qwen3.6-35B-A3B
qwen3.6-35b-a3b
qwen
Q4_K_M
GGUF
moe_partial_cpu_expert_offload
{ "gpu": "NVIDIA GeForce RTX 4060 Ti", "vram_gb": 8, "system_ram_gb": 32, "ram_spec": "DDR5-6000 CL36", "cpu": "AMD Ryzen 5 7600X", "platform": "WSL2 Ubuntu 26.04 on Windows 11" }
{ "engine": "llama-server (llama.cpp b9049-2496f9c14)", "context_length": 32768, "gpu_offload": "max", "expert_offload_to_cpu": true, "kv_cache_quant_k": "q8_0", "kv_cache_quant_v": "q8_0", "flash_attention": true, "thinking_disabled": false, "temperature": 0.7, "top_p": null }
{ "tok_per_sec": 35.36, "ttft_seconds": 1.44, "tokens_generated": 1024, "vram_used_gb": null, "ram_used_gb_peak": null, "prompt_id": "moe-explainer-500w" }
{ "quality_5pt": null, "hit_length_target": null, "notes": "ncmoe=30 (10 layers experts on GPU, 30 on CPU). ngl=999, threads=6. Prompt: 34.72 tok/s. Total: 30.4s. Best overall config. Hybrid SSM+attention means KV cache only scales for 10/40 attention layers — 32K is nearly free. 4.8x faster than full offload. Re...
{ "date": "2026-05-07T00:00:00", "run_type": "cold_server" }
Qwen/Qwen3.6-35B-A3B
qwen3.6-35b-a3b
qwen
Q4_K_M
GGUF
moe_partial_cpu_expert_offload
{ "gpu": "NVIDIA GeForce RTX 4060 Ti", "vram_gb": 8, "system_ram_gb": 32, "ram_spec": "DDR5-6000 CL36", "cpu": "AMD Ryzen 5 7600X", "platform": "WSL2 Ubuntu 26.04 on Windows 11" }
{ "engine": "llama-server (llama.cpp b9049-2496f9c14)", "context_length": 65536, "gpu_offload": "max", "expert_offload_to_cpu": true, "kv_cache_quant_k": "q8_0", "kv_cache_quant_v": "q8_0", "flash_attention": true, "thinking_disabled": false, "temperature": 0.7, "top_p": null }
{ "tok_per_sec": 17.41, "ttft_seconds": 2.83, "tokens_generated": 1024, "vram_used_gb": null, "ram_used_gb_peak": null, "prompt_id": "moe-explainer-500w" }
{ "quality_5pt": null, "hit_length_target": null, "notes": "ncmoe=30 (10 layers experts on GPU, 30 on CPU). ngl=999, threads=6. Prompt: 17.70 tok/s. Total: 61.7s. VRAM cliff at 65K — KV cache for 10 attention layers at q8_0 grows to ~680 MiB, pushing total GPU past ~7 GB. Same ~50% throughput drop as ncmoe 28 at ...
{ "date": "2026-05-07T00:00:00", "run_type": "cold_server" }

Windows RTX 4060 Ti 8GB — MoE Expert Offload Bench (2026-05)

Practitioner test of MoE expert offload on Qwen3.6-35B-A3B with a 32GB-RAM consumer rig. Two regimes tested: full offload (all experts → CPU) and partial offload (-ncmoe N sweep via llama-server from source).

TL;DR

  • Full offload (all experts → CPU): RAM-ceilinged at 7.40 tok/sec on 32GB. Recipe assumes 64GB.
  • Partial offload (-ncmoe 30, 10 layers' experts on GPU): 35.36 tok/sec at 32K context. 4.8x faster.
  • The 8GB VRAM cliff is sharp — crossing ~7 GB total GPU usage halves throughput instantly.

Hardware

Component Spec
GPU RTX 4060 Ti 8GB
System RAM 32GB DDR5-6000 dual-channel (EXPO confirmed at 6000 MT/s)
CPU AMD Ryzen 5 7600X
Platform Windows 11 (full offload: LM Studio native; partial offload: WSL2 Ubuntu 26.04)

Data files

File Regime Runs Date
data/bench-2026-05-06.jsonl Full offload (LM Studio, -ncmoe 99 equivalent) 3 2026-05-06
data/bench-2026-05-07.jsonl Partial offload (-ncmoe sweep via llama-server) 5 2026-05-07

Methodology

Same MoE-explainer prompt across all runs (~500-word target).

Full offload runs (2026-05-06)

  • Engine: LM Studio
  • "Force Model Expert Weights onto CPU" → ON (equivalent to --n-cpu-moe 99)
  • KV cache: q8_0. Flash attention: ON. Thinking: disabled.
  • GPU offload: max. Temperature: 0.7, Top-p: 0.9.

Partial offload runs (2026-05-07)

  • Engine: llama-server (llama.cpp b9049-2496f9c14, ggml 0.11.0) built from source inside WSL2
  • CUDA 13.2.1, compute capability 8.9
  • Flags: -ngl 999 -ncmoe N -fa on --cache-type-k q8_0 --cache-type-v q8_0 -t 6
  • Thinking: ON (Qwen3.6 default — all tokens went to reasoning_content)
  • Temperature: 0.7. Max tokens: 1024. Cold server per config.
  • Variable: -ncmoe (32, 30, 28) and -c (16384, 32768, 65536)

Findings — Full offload

  • Q4_K_S @ 16K: 7.40 tok/sec
  • Q4_K_S @ 32K: 7.43 tok/sec (flat — disk-paging bottleneck, not KV-scan cost)
  • IQ4_XS @ 16K: ~7.40 tok/sec (same speed — quant size is not the lever)

Bottleneck is system RAM size. 32GB system pages expert weights to NVMe during inference.

Findings — Partial offload (-ncmoe sweep)

ncmoe Context Experts on GPU Decode tok/sec Notes
32 16K 8 layers 29.70 Clean, no VRAM pressure
30 16K 10 layers 32.00 Sweet spot at 16K
28 16K 12 layers 16.33 VRAM cliff — PCIe page faults
30 32K 10 layers 35.36 Best overall config
30 65K 10 layers 17.41 VRAM cliff from KV cache growth

Key insights

  1. 8GB VRAM cliff is sharp, not gradual. Crossing ~7 GB total GPU usage triggers CUDA page faults through PCIe. Throughput drops ~50% instantly.
  2. Hybrid SSM+attention makes context scaling cheap. Only 10/40 layers use attention (full_attention_interval=4). KV cache is 1/4 what a pure-attention model needs. 32K costs ~170 MiB extra over 16K.
  3. Optimal config: -ncmoe 30 -c 32768 — 35.36 tok/sec, 4.8x faster than full offload.

Implication

On 8GB VRAM + 32GB RAM, partial expert offload via -ncmoe is the correct approach — not full offload. Building llama.cpp from source gives access to the -ncmoe N flag that LM Studio's slider doesn't expose cleanly. The sweet spot is hardware-specific: find the highest number of expert layers on GPU that fits within ~87% of VRAM.

See also

Downloads last month
41