evan/nix
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

chore(llama-swap): update Genesis to v7.69, add cliff 2 optimizations

Browse Source 
- Bump Genesis pin from fc89395 to 2db18df (v7.69)
- Add PN32 GDN chunked prefill and PN34 workspace lock relax env vars
- Replace patch_workspace_lock_disable with patch_inputs_embeds_optional
- Remove setup-time PN25/PN30 patches (folded into v7.69 natively)
- Switch patch base URL to v7.69-cliff2-test branch
- Lower GPU memory utilization to 0.93 for long-text variant
- Remove python3 from preflight check prerequisites
- Add printing service to lin-va-thinkpad
This commit is contained in:
Evan Reichard
2026-05-02 13:48:41 -04:00
parent 40114f438f
commit 8d45977154
3 changed files with 31 additions and 55 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
modules/nixos/services/llama-swap/config.nix
View File

@@ -98,9 +98,9 @@ in
};
# https://github.com/noonghunna/club-3090/tree/master/models/qwen3.6-27b/vllm
# Synced from: club-3090 ae4846f (2026-05-02) — docker-compose.long-text.yml
# Synced from: club-3090 f6613c8 (2026-05-02) — docker-compose.long-text.yml
# Long-text variant - 180K context, text-only (no vision)
# TurboQuant 3-bit KV + MTP n=3 + Genesis v7.65 full PROD env set
# TurboQuant 3-bit KV + MTP n=3 + Genesis v7.69 + Cliff 2 closure recipe
"vllm-qwen3.6-27b-long-text" = {
name = "vLLM Qwen3.6 (27B) - Long Text";
macros.ctx = "180000";
@@ -110,8 +110,8 @@ in
vllmCmd = ''
set -e; pip install xxhash pandas scipy -q;
python3 -m vllm._genesis.patches.apply_all;
python3 /patches/patch_workspace_lock_disable.py;
python3 /patches/patch_tolist_cudagraph.py;
python3 /patches/patch_inputs_embeds_optional.py;
python3 /patches/patch_timings_07351e088.py;
exec vllm serve
--served-model-name ''${MODEL_ID}
@@ -120,7 +120,7 @@ in
--dtype float16
--tensor-parallel-size 1
--max-model-len ''${ctx}
--gpu-memory-utilization 0.95
--gpu-memory-utilization 0.93
--max-num-seqs 1
--max-num-batched-tokens 4128
--kv-cache-dtype turboquant_3bit_nc
@@ -184,6 +184,11 @@ in
-e GENESIS_ENABLE_P100=1 \
-e GENESIS_ENABLE_P101=1 \
-e GENESIS_ENABLE_P103=1 \
-e GENESIS_ENABLE_PN32_GDN_CHUNKED_PREFILL=1 \
-e GENESIS_PN32_GDN_CHUNK_SIZE=8192 \
-e GENESIS_PN32_GDN_CHUNK_THRESHOLD=16384 \
-e GENESIS_FLA_FWD_H_MAX_T=16384 \
-e GENESIS_ENABLE_PN34_WORKSPACE_LOCK_RELAX=1 \
-e GENESIS_ENABLE_PN8_MTP_DRAFT_ONLINE_QUANT=1 \
-e GENESIS_ENABLE_PN9_INDEPENDENT_DRAFTER_ATTN=1 \
-e GENESIS_ENABLE_PN11_GDN_AB_CONTIGUOUS=1 \
@@ -214,7 +219,7 @@ in
-v /mnt/ssd/vLLM/Cache/triton:/root/.triton/cache \
-v /mnt/ssd/vLLM/Patches/genesis/vllm/_genesis:/usr/local/lib/python3.12/dist-packages/vllm/_genesis:ro \
-v /mnt/ssd/vLLM/Patches/patch_tolist_cudagraph.py:/patches/patch_tolist_cudagraph.py:ro \
-v /mnt/ssd/vLLM/Patches/patch_workspace_lock_disable.py:/patches/patch_workspace_lock_disable.py:ro \
-v /mnt/ssd/vLLM/Patches/patch_inputs_embeds_optional.py:/patches/patch_inputs_embeds_optional.py:ro \
-v /mnt/ssd/vLLM/Patches/patch_timings_07351e088.py:/patches/patch_timings_07351e088.py:ro \
-p ''${PORT}:8000 \
--entrypoint /bin/bash \
@@ -232,9 +237,9 @@ in
};
# https://github.com/noonghunna/club-3090/tree/master/models/qwen3.6-27b/vllm
# Synced from: club-3090 ae4846f (2026-05-02) — docker-compose.long-vision.yml
# Synced from: club-3090 f6613c8 (2026-05-02) — docker-compose.long-vision.yml
# Long-vision variant - 145K context with vision tower active
# TurboQuant 3-bit KV + MTP n=3 + Genesis v7.65 full PROD env set
# TurboQuant 3-bit KV + MTP n=3 + Genesis v7.69 + Cliff 2 env vars (mem-util kept at 0.95)
"vllm-qwen3.6-27b-long-vision" = {
name = "vLLM Qwen3.6 (27B) - Long Vision";
macros.ctx = "145000";
@@ -244,7 +249,6 @@ in
vllmCmd = ''
set -e; pip install xxhash pandas scipy -q;
python3 -m vllm._genesis.patches.apply_all;
python3 /patches/patch_workspace_lock_disable.py;
python3 /patches/patch_tolist_cudagraph.py;
python3 /patches/patch_timings_07351e088.py;
exec vllm serve
@@ -317,6 +321,11 @@ in
-e GENESIS_ENABLE_P100=1 \
-e GENESIS_ENABLE_P101=1 \
-e GENESIS_ENABLE_P103=1 \
-e GENESIS_ENABLE_PN32_GDN_CHUNKED_PREFILL=1 \
-e GENESIS_PN32_GDN_CHUNK_SIZE=8192 \
-e GENESIS_PN32_GDN_CHUNK_THRESHOLD=16384 \
-e GENESIS_FLA_FWD_H_MAX_T=16384 \
-e GENESIS_ENABLE_PN34_WORKSPACE_LOCK_RELAX=1 \
-e GENESIS_ENABLE_PN8_MTP_DRAFT_ONLINE_QUANT=1 \
-e GENESIS_ENABLE_PN9_INDEPENDENT_DRAFTER_ATTN=1 \
-e GENESIS_ENABLE_PN11_GDN_AB_CONTIGUOUS=1 \
@@ -347,7 +356,6 @@ in
-v /mnt/ssd/vLLM/Cache/triton:/root/.triton/cache \
-v /mnt/ssd/vLLM/Patches/genesis/vllm/_genesis:/usr/local/lib/python3.12/dist-packages/vllm/_genesis:ro \
-v /mnt/ssd/vLLM/Patches/patch_tolist_cudagraph.py:/patches/patch_tolist_cudagraph.py:ro \
-v /mnt/ssd/vLLM/Patches/patch_workspace_lock_disable.py:/patches/patch_workspace_lock_disable.py:ro \
-v /mnt/ssd/vLLM/Patches/patch_timings_07351e088.py:/patches/patch_timings_07351e088.py:ro \
-p ''${PORT}:8000 \
--entrypoint /bin/bash \

59
modules/nixos/services/llama-swap/setup-qwen36-vllm.sh
View File

@@ -7,7 +7,7 @@
#
# Idempotent — safe to re-run; skips steps already completed.
#
# Prerequisites: git (with git-lfs), docker, python3
# Prerequisites: git (with git-lfs), docker
set -euo pipefail
@@ -18,9 +18,10 @@ CACHE_DIR="/mnt/ssd/vLLM/Cache"
GENESIS_DIR="${PATCHES_DIR}/genesis"
# Pin Genesis to the validated commit (bump requires re-testing all composes)
GENESIS_PIN="${GENESIS_PIN:-fc89395}"
GENESIS_PIN="${GENESIS_PIN:-2db18df}"
TOLIST_PATCH="${PATCHES_DIR}/patch_tolist_cudagraph.py"
INPUTS_EMBEDS_PATCH="${PATCHES_DIR}/patch_inputs_embeds_optional.py"
WORKSPACE_LOCK_PATCH="${PATCHES_DIR}/patch_workspace_lock_disable.py"
PN25_REGISTER_PATCH="${PATCHES_DIR}/patch_pn25_genesis_register_fix.py"
PN30_DST_PATCH="${PATCHES_DIR}/patch_pn30_dst_shaped_temp_fix.py"
@@ -28,11 +29,11 @@ PR40798_PATCH="${PATCHES_DIR}/patch_pr40798_workspace.py"
TIMINGS_PATCH="${PATCHES_DIR}/patch_timings_07351e088.py"
TIMINGS_PATCH_URL="${TIMINGS_PATCH_URL:-https://gitea.va.reichard.io/evan/nix/raw/branch/master/modules/nixos/services/llama-swap/patches/patch_timings_07351e088.py}"
# Base URL for sidecar patches (club-3090 repo)
PATCH_BASE_URL="https://raw.githubusercontent.com/noonghunna/club-3090/master/models/qwen3.6-27b/vllm/patches"
# Base URL for sidecar patches (club-3090 repo, v7.69-cliff2-test branch)
PATCH_BASE_URL="https://raw.githubusercontent.com/noonghunna/club-3090/v7.69-cliff2-test/models/qwen3.6-27b/vllm/patches"
# ---------- Preflight Checks ----------
for cmd in git git-lfs curl python3; do
for cmd in git git-lfs curl; do
if ! command -v "$cmd" &>/dev/null; then
echo "ERROR: '$cmd' not found in PATH." >&2
exit 1
@@ -87,47 +88,12 @@ download_patch() {
}
download_patch "${TOLIST_PATCH}"
download_patch "${INPUTS_EMBEDS_PATCH}"
download_patch "${WORKSPACE_LOCK_PATCH}"
download_patch "${PN25_REGISTER_PATCH}"
download_patch "${PN30_DST_PATCH}"
download_patch "${PR40798_PATCH}"
# ---------- Apply Setup-Time Genesis Source Patches ----------
# These modify the Genesis checkout in-place. The Genesis tree is mounted
# read-only into the container, so these MUST run at setup time, not boot.
# The setup-time patches use hardcoded relative paths like
# "models/qwen3.6-27b/vllm/patches/genesis/vllm/_genesis/..."
# rooted at the club-3090 repo root. Our Genesis clone lives at
# ${PATCHES_DIR}/genesis, so we create a temporary symlink tree
# so the relative paths resolve correctly.
PATCH_WORKDIR="$(mktemp -d)"
mkdir -p "${PATCH_WORKDIR}/models/qwen3.6-27b/vllm/patches"
ln -sfn "${GENESIS_DIR}" "${PATCH_WORKDIR}/models/qwen3.6-27b/vllm/patches/genesis"
# PN25 Worker-Spawn Registration Fix
# Registers the PN25 opaque op at module import time instead of inside
# the compiled forward path. Required for TP=1 spawned workers.
if [ -f "${PN25_REGISTER_PATCH}" ]; then
echo "Applying PN25 genesis register fix to Genesis tree..."
(cd "${PATCH_WORKDIR}" && python3 "${PN25_REGISTER_PATCH}") || {
echo "WARN: PN25 register fix did not apply cleanly. PN25 may not work in workers." >&2
}
fi
# PN30 DS Conv-State Dst-Shaped Temp Fix
# Corrects DS layout corruption in PN30's speculative decode path by
# building a destination-shaped temp instead of compacting source tail.
if [ -f "${PN30_DST_PATCH}" ]; then
echo "Applying PN30 dst-shaped temp fix to Genesis tree..."
(cd "${PATCH_WORKDIR}" && python3 "${PN30_DST_PATCH}") || {
echo "WARN: PN30 dst-shaped temp fix did not apply cleanly. DS layout may not work." >&2
}
fi
# Clean Up Symlink Workdir
rm -rf "${PATCH_WORKDIR}"
# ---------- Download Timing Patch ----------
tmp_timings_patch="$(mktemp)"
trap 'rm -f "${tmp_timings_patch}"' EXIT
@@ -158,11 +124,12 @@ echo " ├── Cache/"
echo " │ ├── torch_compile/ (torch.compile cache)"
echo " │ └── triton/ (Triton kernel cache)"
echo " └── Patches/"
echo " ├── genesis/ (Genesis v7.65 @ ${GENESIS_PIN})"
echo " │ └── vllm/_genesis/ (mounted into container, PN25+PN30 applied)"
echo " ├── genesis/ (Genesis v7.69 @ ${GENESIS_PIN})"
echo " │ └── vllm/_genesis/ (mounted into container; PN25+PN30+PN34 native)"
echo " ├── patch_tolist_cudagraph.py (boot-time: cudagraph capture fix)"
echo " ├── patch_workspace_lock_disable.py (boot-time: v0.20 WorkspaceManager lock workaround)"
echo " ├── patch_pn25_genesis_register_fix.py (setup-time: applied to Genesis tree)"
echo " ├── patch_pn30_dst_shaped_temp_fix.py (setup-time: applied to Genesis tree)"
echo " ├── patch_inputs_embeds_optional.py (boot-time: vllm#35975 backport, text-only models)"
echo " ├── patch_workspace_lock_disable.py (rollback: superseded by PN34 in v7.69)"
echo " ├── patch_pn25_genesis_register_fix.py (rollback: folded into v7.69 natively)"
echo " ├── patch_pn30_dst_shaped_temp_fix.py (rollback: folded into v7.69 natively)"
echo " ├── patch_pr40798_workspace.py (PR40798 workspace fix)"
echo " └── patch_timings_07351e088.py (boot-time: llama.cpp-compatible timings)"