docs: Qwen3.8 spec-matrix benchmark page; add matrix tools; final env n_max=5
Recording the 54-config matrix, the realistic-prompt correction (n_max=5 not 10), the ubatch findings (256 is the safe/fast point), and the premature-stop sampling-variance investigation.
This commit is contained in:
Frank
@@ -52,6 +52,7 @@ MTP-ONLY draft, port 7770]
|
||||
- [[Running Qwen3.6 35B A3B FP8 on Dankserver]] — MoE model (35B total, 3B active) on RTX 3090
|
||||
- [[Running Qwen3.6 27B FP8 on Dankclaw]] — Dense model (27B) on Strix Halo APU
|
||||
- [[frank-default-large-benchmark]] — backend/spec benchmark; chose MTP sidecar (2026-08-14) over DFlash (parallel premature-EOS bug)
|
||||
- [[qwen3.8-spec-matrix]] — Qwen3.8-27B UD-Q8_K_XL bundled-MTP spec matrix; winner draft-mtp n_max=5
|
||||
- [[frank-default-small-benchmark]] — ncmoe and batch/ubatch sweep that drove the Dankserver state
|
||||
- [[dankserver-vram-growth]] — VRAM/RAM growth investigation behind the recurring cuBLAS crash on frank-default-small
|
||||
- [[Dankclaw Improvements]] — researched enhancements: self-hosted agentic tools, runtime tuning, server/OS settings
|
||||
|
||||
@@ -0,0 +1,86 @@
|
||||
# Qwen3.8-27B UD-Q8_K_XL Speculative-Decoding Matrix
|
||||
|
||||
This page records the speculative-decoding benchmark and the production
|
||||
tuning for the unsloth Qwen3.8-27B `UD-Q8_K_XL` service on Dankclaw
|
||||
(`frank-default-large`).
|
||||
|
||||
## Result (applied)
|
||||
|
||||
- Model: `Qwen3.8-27B-UD-Q8_K_XL.gguf` (29.29 GiB, bundled MTP).
|
||||
- Spec: `draft-mtp`, `--spec-draft-n-max 5`, no `--spec-draft-p-min`.
|
||||
- Batch 8192 / ubatch 256, parallel 2, ctx 524288 (262144 per slot).
|
||||
- Realistic-prompt generation: ~13.3 t/s single, ~8 t/s per slot under
|
||||
2 concurrent streams. No premature stop.
|
||||
|
||||
## How MTP works on Qwen3.8
|
||||
|
||||
The unsloth `UD-*` GGUFs keep the MTP (Multi-Token Prediction) head inside
|
||||
the model file. llama.cpp b10411+ auto-detects it and creates the draft
|
||||
context from the target model. Do not pass `--spec-draft-model`.
|
||||
|
||||
The stock llama.cpp build must be b10411 or newer. The 2026-08-14 rebuild
|
||||
on Dankclaw is b10433 (`9b05354ec`) with `GGML_HIP` + `gfx1151` + Vulkan.
|
||||
|
||||
## Benchmark Method
|
||||
|
||||
The matrix ran 54 configurations: 3 spec types (`draft-mtp,ngram-mod`,
|
||||
`draft-mtp`, `ngram-mod`) x `--spec-draft-n-max` 2..10 x
|
||||
`--spec-draft-p-min` {off, 0.8}. Each config loaded `llama-server`,
|
||||
then ran 3 generations. The tool is `scripts/spec-matrix-bench.sh`;
|
||||
analysis is `scripts/analyze-spec-matrix.py`.
|
||||
|
||||
## Matrix Caveat: Benchmark Prompt
|
||||
|
||||
The matrix used a repetitive 1100-token prompt
|
||||
(`scripts/bench_live.sh` style). MTP drafts that repeated sentence almost
|
||||
perfectly, which inflated the numbers for long drafts. The matrix reported
|
||||
`draft-mtp` n_max=10 at ~30 t/s. A realistic story prompt gave very
|
||||
different results:
|
||||
|
||||
| n_max | median t/s (realistic) |
|
||||
| ----- | ----------------------- |
|
||||
| 4 | 12.5 |
|
||||
| 5 | 13.3 (peak) |
|
||||
| 6 | 10.6 |
|
||||
| 7 | 10.1 |
|
||||
| 10 | 8.6 |
|
||||
| none (baseline) | 7.2 |
|
||||
|
||||
Long drafts collapse to ~18% acceptance on real writing. n_max=5 balances
|
||||
acceptance against draft cost.
|
||||
|
||||
## Premature Stop Investigation
|
||||
|
||||
Two failure patterns appeared during early production tests:
|
||||
|
||||
1. Speed collapse. `ubatch >= 512` reduced MTP acceptance to ~25% and
|
||||
generation to ~9 t/s. `ubatch 128` with `batch >= 2048` triggered a
|
||||
premature instant-EOS under 2 concurrent slots. `ubatch 256` avoids
|
||||
both: ~30 t/s single, concurrent-safe.
|
||||
2. Apparent premature-EOS. The repetitive benchmark prompt with `temp 1.0`
|
||||
sometimes stops after 1-2 tokens because the model legitimately emits an
|
||||
EOS on a wall of repeated text. This is sampling variance, not a
|
||||
corruption bug. Realistic prompts never stop early.
|
||||
|
||||
## Production Configuration
|
||||
|
||||
`scripts/frank-default-large.env`:
|
||||
|
||||
```ini
|
||||
SPEC_TYPE=draft-mtp
|
||||
SPEC_DRAFT_N_MAX=5
|
||||
SPEC_DRAFT_P_MIN=
|
||||
BATCH_SIZE=8192
|
||||
UBATCH_SIZE=256
|
||||
CTX_SIZE=524288
|
||||
PARALLEL=2
|
||||
```
|
||||
|
||||
Window: 524288 total / 2 = 262144 per slot, matching the pi
|
||||
`models.json` contextWindow for `frank-default-large`.
|
||||
|
||||
## See Also
|
||||
|
||||
- [[frank-default-large-benchmark]]
|
||||
- [[ling-3.0-flash-trial]]
|
||||
- [[deployment]]
|
||||
@@ -0,0 +1,70 @@
|
||||
#!/usr/bin/env python3
|
||||
"""Analyze the spec-matrix TSV: per-config median gen t/s, acceptance,
|
||||
sorted, with the top configs and per-spec-type bests."""
|
||||
import sys, statistics, collections
|
||||
|
||||
path = sys.argv[1] if len(sys.argv) > 1 else "/tmp/spec_matrix_results.tsv"
|
||||
rows = []
|
||||
with open(path) as f:
|
||||
header = f.readline().strip().split("\t")
|
||||
for line in f:
|
||||
p = line.rstrip("\n").split("\t")
|
||||
if len(p) < 9:
|
||||
continue
|
||||
spec, n_max, p_min, run, tps, pps, acc, dn, stop = p[:9]
|
||||
try:
|
||||
tps = float(tps)
|
||||
except ValueError:
|
||||
continue
|
||||
rows.append(dict(spec=spec, n_max=n_max, p_min=p_min, run=run,
|
||||
tps=tps, pps=float(pps or 0), acc=acc, dn=dn, stop=stop))
|
||||
|
||||
by_cfg = collections.defaultdict(list)
|
||||
for r in rows:
|
||||
by_cfg[(r["spec"], r["n_max"], r["p_min"])].append(r)
|
||||
|
||||
print("%-22s %5s %5s %8s %8s %10s" % ("spec", "n_max", "p_min", "med_tps", "med_pps", "accept"))
|
||||
summary = []
|
||||
for (spec, n_max, p_min), runs in by_cfg.items():
|
||||
tps_list = [r["tps"] for r in runs]
|
||||
med = statistics.median(tps_list)
|
||||
pps_list = [r["pps"] for r in runs]
|
||||
accs = [r["acc"] for r in runs if r["acc"] not in ("", "None")]
|
||||
acc_s = ""
|
||||
if accs:
|
||||
try:
|
||||
nums = [float(a) for a in accs]
|
||||
acc_s = "%.3f" % (sum(nums)/len(nums))
|
||||
except ValueError:
|
||||
pass
|
||||
summary.append((med, spec, n_max, p_min,
|
||||
statistics.median(pps_list), acc_s, len(runs)))
|
||||
|
||||
summary.sort(reverse=True)
|
||||
for med, spec, n_max, p_min, pps, acc, n in summary:
|
||||
print("%-22s %5s %5s %8.2f %8.2f %10s" % (spec, n_max, p_min, med, pps, acc))
|
||||
|
||||
print("\n=== TOP 12 CONFIGS ===")
|
||||
for med, spec, n_max, p_min, pps, acc, n in summary[:12]:
|
||||
print(" %5.2f t/s spec=%-22s n_max=%2s p_min=%s" % (med, spec, n_max, p_min))
|
||||
|
||||
print("\n=== BEST PER SPEC TYPE ===")
|
||||
best = {}
|
||||
for med, spec, n_max, p_min, pps, acc, n in summary:
|
||||
key = spec
|
||||
if key not in best:
|
||||
best[key] = (med, n_max, p_min, pps, acc)
|
||||
for k, v in best.items():
|
||||
print(" %-22s %5.2f t/s n_max=%s p_min=%s (pps=%s)" % (k, v[0], v[1], v[2], v[3]))
|
||||
|
||||
print("\n=== p_min 0.8 vs off (per spec/n_max, med delta) ===")
|
||||
from itertools import product
|
||||
for spec in sorted(set(r["spec"] for r in rows)):
|
||||
print(" spec:", spec)
|
||||
for n_max in sorted(set(int(r["n_max"]) for r in rows if r["spec"] == spec)):
|
||||
on = [r["tps"] for r in rows if r["spec"] == spec and r["n_max"] == str(n_max) and r["p_min"] == "0.8"]
|
||||
off = [r["tps"] for r in rows if r["spec"] == spec and r["n_max"] == str(n_max) and r["p_min"] == "none"]
|
||||
if on and off:
|
||||
d = statistics.median(on) - statistics.median(off)
|
||||
print(" n_max=%-2s on=%.2f off=%.2f delta=%+.2f" % (n_max,
|
||||
statistics.median(on), statistics.median(off), d))
|
||||
@@ -0,0 +1,78 @@
|
||||
#!/usr/bin/env bash
|
||||
# bench-spec-matrix.sh - Qwen3.8-27B UD-Q8_K_XL speculative-decode matrix.
|
||||
# Matrix: spec_type {draft-mtp,ngram-mod | draft-mtp | ngram-mod}
|
||||
# x --spec-draft-n-max {2..10} x --spec-draft-p-min {off, 0.8}
|
||||
# = 54 configs. Each: start llama-server, wait ready, 3 generations of a
|
||||
# 1100-token prompt (n_predict 200, default sampling), record timings.
|
||||
set -u
|
||||
|
||||
MODEL=/home/dank/models/Qwen3.8-27B-UD-Q8_K_XL.gguf
|
||||
LLAMA_SERVER=/home/dank/llama.cpp/build/bin/llama-server
|
||||
PORT=7890
|
||||
DEVICE=ROCm0
|
||||
OUT=/tmp/spec_matrix_results.tsv
|
||||
RUNLOG=/tmp/spec_matrix_run.log
|
||||
PROMPT=$(python3 -c "print(\"The fleet of inference servers must remain available at all times. Operators monitor memory, throughput, and queue depth. \"*50)")
|
||||
N_OUT=200
|
||||
RUNS=3
|
||||
|
||||
: > "$OUT"
|
||||
printf 'spec_type\tn_max\tp_min\trun\ttps\tpps\tdraft_acc\tdraft_n\tstop\n' >> "$OUT"
|
||||
|
||||
cfg=0
|
||||
for spec in "draft-mtp,ngram-mod" "draft-mtp" "ngram-mod"; do
|
||||
for n_max in 2 3 4 5 6 7 8 9 10; do
|
||||
for p_min in none 0.8; do
|
||||
cfg=$((cfg+1))
|
||||
echo "=== CONFIG $cfg/54: spec=$spec n_max=$n_max p_min=$p_min $(date +%T) ===" | tee -a "$RUNLOG"
|
||||
ARGS=(--model "$MODEL" --spec-type "$spec" --spec-draft-n-max "$n_max")
|
||||
# Qwen3.8 bundles its MTP drafter inside the target GGUF; no separate
|
||||
# --spec-draft-model file (llama.cpp auto-detects, b10411+).
|
||||
if [[ "$p_min" != none ]]; then
|
||||
ARGS+=(--spec-draft-p-min "$p_min")
|
||||
fi
|
||||
ARGS+=(--device "$DEVICE" --gpu-layers all --cache-type-k q8_0 --cache-type-v f16
|
||||
--ctx-size 2048 --parallel 1 --batch-size 512 --ubatch-size 128
|
||||
--temp 1.0 --top-k 20 --top-p 0.95 --min-p 0.0
|
||||
--host 127.0.0.1 --port "$PORT")
|
||||
"$LLAMA_SERVER" "${ARGS[@]}" > /tmp/matrix_server.log 2>&1 &
|
||||
SPID=$!
|
||||
ready=0
|
||||
for i in $(seq 1 150); do
|
||||
if curl -sf "http://127.0.0.1:$PORT/health" > /dev/null 2>&1; then ready=1; break; fi
|
||||
if ! kill -0 "$SPID" 2>/dev/null; then break; fi
|
||||
sleep 2
|
||||
done
|
||||
if [[ "$ready" != 1 ]]; then
|
||||
echo "NOT_READY: spec=$spec n_max=$n_max p_min=$p_min" | tee -a "$RUNLOG"
|
||||
tail -4 /tmp/matrix_server.log >> "$RUNLOG"
|
||||
kill "$SPID" 2>/dev/null; wait "$SPID" 2>/dev/null
|
||||
continue
|
||||
fi
|
||||
for r in 1 2 3; do
|
||||
curl -s -m 300 -X POST "http://127.0.0.1:$PORT/completion" -H "Content-Type: application/json" \
|
||||
-d "{\"prompt\": \"$PROMPT\", \"n_predict\": $N_OUT, \"stream\": false, \"cache_prompt\": false}" \
|
||||
| python3 -c "
|
||||
import json,sys
|
||||
try:
|
||||
d=json.load(sys.stdin)
|
||||
t=d.get('timings',{})
|
||||
row=[\"$spec\",\"$n_max\",\"$p_min\",\"$r\",
|
||||
str(round(t.get('predicted_per_second',0),2)),
|
||||
str(round(t.get('prompt_per_second',0),2)),
|
||||
str(t.get('draft_n_accepted','')),
|
||||
str(t.get('draft_n','')),
|
||||
str(d.get('stop_type',''))]
|
||||
with open(\"$OUT\",\"a\") as f:
|
||||
f.write('\t'.join(row)+'\n')
|
||||
except Exception as e:
|
||||
with open(\"$OUT\",\"a\") as f:
|
||||
f.write('\t'.join([\"$spec\",\"$n_max\",\"$p_min\",\"$r\",\"ERR\",str(e),\"\",\"\",\"\"])+'\n')
|
||||
"
|
||||
done
|
||||
kill "$SPID" 2>/dev/null; wait "$SPID" 2>/dev/null
|
||||
sleep 3
|
||||
done
|
||||
done
|
||||
done
|
||||
echo "=== MATRIX DONE $(date) ===" | tee -a "$RUNLOG"
|
||||
Reference in New Issue
Block a user