Llama 3.3 70B · llama.cpp · Q4_K_M

status: done

Configuration

Modelmeta-llama/Llama-3.3-70B-Instruct
CompanyMeta
FamilyLlama
Parameters70B (dense)
Enginellama.cpp
Quant / precisionQ4_K_M
Why this quantGGUF Q4_K_M from unsloth (trusted quantizer) — widest llama.cpp coverage, strong size/quality balance.
Downloadunsloth/Llama-3.3-70B-Instruct-GGUF
Context window65536
Input modalitiestext

Measured results

Prefill tok/s--num-prompts 1000 --max-seconds 900 --concurrency 32 --max-tokens 256
Decode tok/s48.7
Peak memory (GB)70.94 (system MemAvailable delta (10s sampling))
Completed2026-06-22 16:01 +08

Full run command

# ghcr.io/ggml-org/llama.cpp:full-cuda (dispatcher → --server). GGUF under /home/gauravmm/models.
docker run --gpus all -p 8081:8081 -v /home/gauravmm/models:/models:ro \
  ghcr.io/ggml-org/llama.cpp:full-cuda \
  --server -m /models/Llama-3.3-70B-Instruct-Q4_K_M.gguf -ngl 99 -c 65536 \
  --parallel 32 -cb --host 0.0.0.0 --port 8081
python3 scripts/bench-serving.py --base-url http://localhost:8081 \
  --model Llama-3.3-70B-Instruct-Q4_K_M.gguf \
  --dataset benchmark_data/ShareGPT_V3_unfiltered_cleaned_split.json \
  --num-prompts 1000 --max-seconds 900 --concurrency 32 --max-tokens 256

llama-3.3-70b Meta Llama Q4_K_M 41-130B conc-32

Notes

The dense-70B tier opens the 41-130B bucket — roughly half the decode of the dense 32Bs, exactly as the parameter count predicts. Meta’s Llama-3.3-70B, unsloth Q4_K_M.

  • Workload: ShareGPT V3, concurrency 32. 273/1000, 6 errors, hit the 15-min cap (drained at 1205 s as the last in-flight batch finished). Loaded in 35 s.
  • Throughput (aggregate, conc 32): prefill 62.2 tok/s, decode 48.7 tok/s. TTFT median 2.0 s, TPOT median 476 ms (≈2.1 tok/s/stream), req throughput 0.23/s.
  • Clean parameter scaling on llama.cpp. A dense 70B fires ~2.2× the weights of the dense 32Bs per token, and decode lands at 48.7 vs the 32Bs’ ~112 — almost exactly the inverse-parameter ratio. At conc 32 the GB10’s ~273 GB/s of memory bandwidth is the wall: 70B of Q4 weights streamed per token caps throughput here. This is the slowest decode of any non-reasoning model so far, and a capability/fit test as much as a speed one.
  • Memory: 70.9 GB — ~40 GB Q4_K_M weights + Llama’s GQA KV at 64K ctx. Fits the 121 GB machine with headroom, unlike the Gemma-31B global-attention cliff (88 GB for half the parameters).
  • Slot-split errors (6): -c 65536 --parallel 32 → 2048 tok/slot. Low count only because so few requests completed before the cap.
  • For higher 70B-class throughput, an FP8/AWQ build on vLLM (batched kernels) would beat llama.cpp here, the same pattern seen across the dense models — but the dense 70B is fundamentally bandwidth-bound on a single Spark.