Qwen3.5 9B · Q4_K_M by unsloth

Raw logs

• kld
• llama_bench
• llama_server
• toolcall
• toolcall_v0_1_1

Full run record (JSON)

{
  "id": "20261021T053418_qwen3.5-9b--unsloth--q4_k_m",
  "quant_id": "qwen3.5-9b--unsloth--q4_k_m",
  "started_at": "2026-10-21T05:34:18Z",
  "finished_at": "2026-10-21T05:41:44Z",
  "status": "partial",
  "machine": {
    "name": "Ghost",
    "gpu": "NVIDIA GeForce RTX 5070",
    "vram_gb": 12,
    "ram_gb": 64,
    "cpu": "AMD Ryzen 7 5800X",
    "os": "Windows 11",
    "driver": "616.92",
    "cuda": "13.4"
  },
  "software": {
    "llama_cpp_build": "b11770 (39da130fc)",
    "lm_eval_version": "0.4.13",
    "harness_version": "0.1.1",
    "python": "3.12.10"
  },
  "placement": "full_gpu",
  "settings": {
    "commands": [
      "C:\\quantdiff\\llama.cpp\\bin\\llama-perplexity.exe -m C:\\quantdiff\\hf-cache\\models--unsloth--Qwen3.5-9B-GGUF\\snapshots\\3885219b6810b007914f3a7950a8d1b469d598a5\\Qwen3.5-9B-Q4_K_M.gguf --kl-divergence-base C:\\quantdiff\\work\\reference\\qwen3.5-9b\\quantdiff-kld-v1.kld --kl-divergence -c 512 -b 512 -ngl 999",
      "C:\\quantdiff\\llama.cpp\\bin\\llama-bench.exe -m C:\\quantdiff\\hf-cache\\models--unsloth--Qwen3.5-9B-GGUF\\snapshots\\3885219b6810b007914f3a7950a8d1b469d598a5\\Qwen3.5-9B-Q4_K_M.gguf -ngl 999 -fa 1 -p 512 -n 128 -r 5 -t 8 -o json",
      "C:\\quantdiff\\llama.cpp\\bin\\llama-server.exe -m C:\\quantdiff\\hf-cache\\models--unsloth--Qwen3.5-9B-GGUF\\snapshots\\3885219b6810b007914f3a7950a8d1b469d598a5\\Qwen3.5-9B-Q4_K_M.gguf --host 127.0.0.1 --port 8089 -c 4096 -np 1 -ngl 999 -fa on --jinja --seed 1234 --no-webui -a quant"
    ],
    "ctx": 4096,
    "batch": 512,
    "threads": 8,
    "flash_attn": true,
    "cache_type_k": "f16",
    "cache_type_v": "f16",
    "seed": 1234,
    "temperature": 0.0,
    "kld_corpus": "quantdiff-kld-v1",
    "kld_corpus_sha256": "ff19568db86d014569a44b6da4c5830fbcd26de59f57c9ece87b45c34d9c5782",
    "kld_chunks": 100,
    "kld_ctx": 512,
    "kld_reference": "Qwen/Qwen3.5-9B -> convert_hf_to_gguf.py --outtype bf16 (cached)",
    "kld_ref_ppl": 7.257,
    "ngl": 999,
    "gpu_info": "NVIDIA GeForce RTX 5070",
    "server_ctx": 4096,
    "server_ngl": 999,
    "toolcall_set": "quantdiff-bfcl-v1",
    "toolcall_set_sha256": "4d0e1667163c96a109965c58242673528c62318322450d9b218a730c09965020",
    "toolcall_chat_template_kwargs": {
      "enable_thinking": false
    },
    "toolcall_parallel_tool_calls": true,
    "toolcall_rerun_at": "2026-10-21T08:12:34Z",
    "toolcall_harness_version": "0.1.2"
  },
  "results": {
    "kld": {
      "mean": 0.046212,
      "max": 20.196701,
      "p99": 0.475209,
      "median": 0.013264,
      "same_top_pct": 92.443,
      "ppl": 7.364546,
      "ppl_ref": 7.22977,
      "delta_p_rms": 6.225
    },
    "speed": {
      "pp_tokens": 512,
      "tg_tokens": 128,
      "pp_tps": 4699.97,
      "pp_stddev": 128.7,
      "tg_tps": 100.97,
      "tg_stddev": 0.18
    },
    "toolcall": {
      "score": 91.5,
      "passed": 183,
      "total": 200,
      "request_errors": 0,
      "by_category": {
        "simple_python": {
          "passed": 58,
          "total": 60,
          "score": 96.7
        },
        "multiple": {
          "passed": 36,
          "total": 40,
          "score": 90.0
        },
        "parallel": {
          "passed": 30,
          "total": 30,
          "score": 100.0
        },
        "parallel_multiple": {
          "passed": 26,
          "total": 30,
          "score": 86.7
        },
        "irrelevance": {
          "passed": 33,
          "total": 40,
          "score": 82.5
        }
      }
    }
  },
  "logs": {
    "kld": "kld.log",
    "llama_bench": "llama_bench.log",
    "llama_server": "llama_server.log",
    "toolcall": "toolcall.jsonl",
    "toolcall_v0_1_1": "toolcall.v0.1.1.jsonl"
  },
  "example": false,
  "notes": "tool-call stage re-run 2026-10-21T08:12:34Z with parallel_tool_calls=true (first pass used llama-server's default, which allows at most one call per turn; old score 91.5%)"
}