dank/serverAppConfigs
1
0
Fork 0
Code Issues 6 Pull Requests Actions Packages Projects Releases Wiki Activity

Configure DeepSeek reasoning and large-model ubatch sweep

Browse Source
This commit is contained in:
Frank
2026-08-25 16:52:45 -07:00
parent 8c8c3f7cd0
commit d6ea71b9a4
7 changed files with 99 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/LLM/deployment.md
+1 -1
View File
@@ -138,7 +138,7 @@ from `peculiar-ragdoll/Qwen-Sharp-Chat-Templates`. See [[dealign-ornith-crack]].
### Dankclaw — frank-default-large
The active service on Dankclaw uses the ROCmFPX llama.cpp runtime to serve
[kingjones777/Qwen3.8-Uncensored-ROCmFP4-STRIX-MTP-GGUF](https://huggingface.co/kingjones777/Qwen3.8-Uncensored-ROCmFP4-STRIX-MTP-GGUF)
[kingjones777/Qwen3.8-27B-Uncensored-ROCmFP4-STRIX-MTP-GGUF](https://huggingface.co/kingjones777/Qwen3.8-27B-Uncensored-ROCmFP4-STRIX-MTP-GGUF)
under the alias `frank-default-large` on port 7770. The main model is
`Qwen3.8-27B-Uncensored-Q6_0_ROCMFPX_AGENT.gguf`; the standalone MTP draft is
`mtp-Qwen3.8-27B-Q4_0.gguf`, and the Q8_0 vision projector is loaded alongside
docs/LLM/index.md
+1
View File
@@ -50,6 +50,7 @@ MTP draft + Turbo4 V cache, 262K ctx / 1 slot]
- [[frank-redo-2026-08-17]] — the 2026-08-17 topology change
- [kingjones Qwen3.8 Uncensored ROCmFP4 STRIX MTP](https://huggingface.co/kingjones777/Qwen3.8-27B-Uncensored-ROCmFP4-STRIX-MTP-GGUF) — active frank-default-large model on Dankclaw; Q6 AGENT main model with standalone MTP
- [[kingjones-strix-mtp]] — the active Q6 AGENT runbook: serving profile, standalone MTP, reasoning + preserve_thinking, 114 GiB TTM/GTT
- [[ubatch-benchmark-large]] — active-model ubatch sweep plan and results
- [[kingjones-q8agent-rocmfpx]] — the retired Q8_0-AGENT investigation and `set_rows` failure
- [[eemin-rocmi4-mtp]] — the former caf-eemin Qwen3.8-27B embedded-MTP model
on Dankclaw (retired)
docs/LLM/ubatch-benchmark-large.md
+59
View File
@@ -0,0 +1,59 @@
---
type: benchmark plan
title: frank-default-large ubatch sweep
description: Standardized ubatch sweep for the active kingjones Qwen3.8 ROCmFPX deployment.
tags:
- llama.cpp
- rocm
- benchmark
- ubatch
timestamp: 2026-09-01
---
# frank-default-large ubatch sweep
This benchmark sweeps the active `frank-default-large` model on Dankclaw using
scratch ports and an 8K-context workload. The production service is stopped
before the sweep and restarted by the benchmark's exit handler.
## Sweep
The default points are `256, 512, 768, 1024, 1280, 1536, 2048, 3072, 4096,
5120, 6144`. This implements the requested 256-step range through 1536 and
1024-step range from 2048 through 6144. Each point runs three 256-token decode
requests by default.
Run on Dankclaw from the repository checkout:
```bash
STOPSVC=1 ./scripts/bench-ubatch-rocmi4.sh
```
The script uses the active Q6 AGENT model, standalone MTP draft, tracked Sharp
template, and `--reasoning-format deepseek`. It reports server timing lines and
per-run generation throughput. Run only when the production service can be
interrupted; the GPU is shared with live workloads.
## Results
Record results below after the sweep, including failures or OOM points.
| ubatch | prefill | generation | notes |
| ---: | ---: | ---: | --- |
| 256 | — | — | pending |
| 512 | — | — | pending |
| 768 | — | — | pending |
| 1024 | — | — | pending |
| 1280 | — | — | pending |
| 1536 | — | — | pending |
| 2048 | — | — | pending |
| 3072 | — | — | pending |
| 4096 | — | — | pending |
| 5120 | — | — | pending |
| 6144 | — | — | pending |
## See also
- [LLM fleet index](index.md)
- [Deployment](deployment.md)
- [Active kingjones runbook](kingjones-strix-mtp.md)
docs/index.md
+1
View File
@@ -116,6 +116,7 @@ The LLM inference fleet on the two hosts. Full index:
### Benchmarks and Sizing
- [[frank-default-large-benchmark]] — backend/spec benchmark behind the old head
- [[ubatch-benchmark-large]] — active-model ubatch sweep plan and results
- [[frank-default-small-benchmark]] — the small-model benchmark suite
- [[Running Qwen3.6 35B A3B FP8 on Dankserver]] — Qwen3.6-35B-A3B hardware notes
- [[Running Qwen3.6 27B FP8 on Dankclaw]] — Qwen3.6-27B sizing on the iGPU
scripts/bench-ubatch-rocmi4.sh
+16 -14
View File
@@ -1,6 +1,6 @@
#!/usr/bin/env bash
# bench-ubatch-rocmi4.sh — sweep ubatch size at 8k context on the
# frank-default-large ROCmI4 embedded-MTP stack (dankclaw, gfx1151).
# frank-default-large ROCmFPX standalone-MTP stack (dankclaw, gfx1151).
#
# Boots one llama-server instance per ubatch on a scratch port with a fixed
# -b and -c 8192, runs a standardized ~75%-of-ctx prose workload, and reads
@@ -26,21 +26,23 @@ if [[ -f "$ENV_FILE" ]]; then
fi
# --- sweep parameters -------------------------------------------------------
: "${UBATCH_MIN:=512}"
: "${UBATCH_MAX:=4096}"
: "${UBATCH_STEP:=512}"
: "${UBATCH_MIN:=256}"
: "${UBATCH_MAX:=6144}"
: "${UBATCH_STEP:=256}"
: "${UBATCH_VALUES:=256 512 768 1024 1280 1536 2048 3072 4096 5120 6144}"
: "${CTX:=8192}" # 8k contexts
: "${BENCH_BATCH:=$CTX}" # fixed batch = ctx, so the full 8k prompt fits
: "${OUT_TOKENS:=256}"
: "${RUNS:=3}" # decode runs per ubatch
: "${PORT_BASE:=7894}" # scratch ports 7894..
: "${STOPSVC:=0}" # stop/restart the live unit for a clean GPU
: "${STOPSVC:=1}" # stop/restart the live unit for a clean GPU
# --- resolved paths (fall back to serve-script defaults) --------------------
: "${LLAMA_BIN:=/home/dank/rocmi4-qwen38/ROCmFPX/build-rocmi4-w4a4/bin/llama-server}"
: "${MODEL_DIR:=/home/dank/models/eemin-Qwen3.8-27B-ROCMI4-MTP}"
: "${MODEL_FILE:=Qwen3.8-27B-Q4_0_ROCMI4.gguf}"
: "${MMPROJ_FILE:=mmproj-BF16.gguf}"
: "${LLAMA_BIN:=/home/dank/ciru-qwen38/ROCmFPX/build-promptforge/bin/llama-server}"
: "${MODEL_DIR:=/home/dank/models/kingjones-Qwen3.8-27B-STRIX-MTP}"
: "${MODEL_FILE:=Qwen3.8-27B-Uncensored-Q6_0_ROCMFPX_AGENT.gguf}"
: "${MTP_DRAFT_FILE:=$MODEL_DIR/mtp-Qwen3.8-27B-Q4_0.gguf}"
: "${MMPROJ_FILE:=mmproj-Qwen3.8-27B-Q8_0.gguf}"
: "${CHAT_TEMPLATE_FILE:=$SCRIPT_DIR/chat-template.jinja}"
: "${ALIAS:=frank-default-large}"
: "${HSA_OVERRIDE_GFX_VERSION:=11.5.1}"
@@ -114,7 +116,7 @@ run_one() {
--model "$MODEL" \
"${MMPROJ_ARGS[@]}" \
--chat-template-file "$CHAT_TEMPLATE_FILE" \
--chat-template-kwargs '{"enable_thinking":true,"preserve_thinking":true}' \
--chat-template-kwargs '{"enable_thinking":true,"preserve_thinking":true,"preserve_reasoning":true}' \
--alias bench-ubatch \
--host 127.0.0.1 --port "$port" \
-dev ROCm0 -ngl 999 -np 1 \
@@ -122,14 +124,14 @@ run_one() {
-c "$CTX" -b "$BENCH_BATCH" -ub "$ubatch" \
-fa on -ctk f16 -ctv f16 -ctxcp 16 \
--cache-ram 8192 --cache-prompt --metrics --no-mmap \
--spec-type draft-mtp --spec-mtp-strict-qwen \
--spec-draft-device ROCm0 --spec-draft-ngl all \
--spec-type draft-mtp --model-draft "$MTP_DRAFT_FILE" \
--spec-draft-device ROCm0 --spec-draft-ngl 999 \
--spec-draft-type-k f16 --spec-draft-type-v f16 \
--spec-draft-n-max "$MTP_N_MAX" --spec-draft-n-min 0 \
--spec-draft-p-min "$MTP_P_MIN" \
--spec-draft-backend-sampling \
--temp 0.7 --top-p 0.95 \
--reasoning on --reasoning-format deepseek --reasoning-budget -1 \
--reasoning "$REASONING" --reasoning-format "$REASONING_FORMAT" --reasoning-budget -1 \
-lv 4 >>"$log" 2>&1 &
local pid=$!
wait_health "$pid" "$port" || { tail -30 "$log"; return 2; }
@@ -207,7 +209,7 @@ main() {
stop_live_service
local idx=0 ubatch port
printf "%-10s %-14s %-10s\n" "UBATCH" "PREFILL" "GEN_TPS"
for ubatch in $(seq "$UBATCH_MIN" "$UBATCH_STEP" "$UBATCH_MAX"); do
for ubatch in $UBATCH_VALUES; do
port=$((PORT_BASE + idx))
run_one "$ubatch" "$port" "$OUT_TOKENS" "$RUNS"
idx=$((idx + 1))
scripts/frank-default-large.env
+6 -5
View File
@@ -22,8 +22,9 @@ MTP_DRAFT_FILE=$MODEL_DIR/mtp-Qwen3.8-27B-Q4_0.gguf
MMPROJ_FILE=mmproj-Qwen3.8-27B-Q8_0.gguf
# Qwen-Sharp chat template (tracked in the repo, staged on the host). The Sharp
# template drives thinking on by default (medium effort) and produces the
# peg-native + deepseek chat/reasoning shape the Pi agent expects.
# template drives thinking on by default (medium effort); the server emits
# reasoning in DeepSeek format so clients receive reasoning_content separately
# instead of literal <think> tags in content.
CHAT_TEMPLATE_FILE=/home/dank/serverAppConfigs/scripts/chat-template.jinja
CHAT_TEMPLATE_KWARGS='{"enable_thinking":true,"preserve_thinking":true,"preserve_reasoning":true}'
@@ -63,10 +64,10 @@ MIN_P=0.0
PRESENCE_PENALTY=1.5
# Reasoning on (Sharp template drives it; preserve_thinking + preserve_reasoning
# keep the reasoning trail in the conversation). reasoning-format none lets the
# template project the reasoning content into reasoning_content.
# keep the reasoning trail in the conversation). DeepSeek format projects the
# reasoning into the API's reasoning_content field instead of content tags.
REASONING=on
REASONING_FORMAT=none
REASONING_FORMAT=deepseek
# Strix Halo unified memory + native gfx1151 detection for this ROCmFPX build.
HSA_OVERRIDE_GFX_VERSION=11.5.1
tests/frank_large_config.bats
+15
View File
@@ -39,6 +39,11 @@ TEMPLATE_FILE="$REPO_ROOT/scripts/chat-template.jinja"
grep -q ': "${CTX_SIZE:=262144}"' "$SERVE_FILE"
}
@test "frank-default-large uses DeepSeek reasoning output" {
grep -q '^REASONING_FORMAT=deepseek$' "$ENV_FILE"
grep -q -- '--reasoning-format "$REASONING_FORMAT"' "$SERVE_FILE"
}
@test "frank-default-large has reasoning and preserved reasoning/thinking enabled" {
grep -q '^REASONING=on$' "$ENV_FILE"
grep -q 'enable_thinking":true' "$ENV_FILE"
@@ -58,6 +63,16 @@ TEMPLATE_FILE="$REPO_ROOT/scripts/chat-template.jinja"
grep -q -- '-ub "$UBATCH_SIZE"' "$SERVE_FILE"
}
@test "large ubatch bench defaults to the active model and requested sweep" {
local bench="$REPO_ROOT/scripts/bench-ubatch-rocmi4.sh"
grep -q ': "${UBATCH_MIN:=256}"' "$bench"
grep -q ': "${UBATCH_MAX:=6144}"' "$bench"
grep -q 'UBATCH_VALUES.*256 512 768 1024 1280 1536 2048 3072 4096 5120 6144' "$bench"
grep -q ': "${STOPSVC:=1}"' "$bench"
grep -q 'MODEL_FILE' "$bench"
grep -q 'REASONING_FORMAT' "$bench"
}
@test "frank-default-large uses the tracked Qwen-Sharp template on :7770" {
grep -q '^CHAT_TEMPLATE_FILE=' "$ENV_FILE"
grep -q -- '--jinja' "$SERVE_FILE"