Frank
780c307354
The first crash-looping boot (51 restarts): serve received the host path /home/dank/models/... but the models tree is mounted at /models inside the container, so transformers failed 'Repo id must be in the form namespace/repo_name' before the engine even built. Same conversion the spec draft already used. Pinned by bats. Root docs index links the new runbook (seam required it) and the work log records the crash-loop plus the 52 G payload reclaim.
207 lines
10 KiB
Bash
Executable File
207 lines
10 KiB
Bash
Executable File
#!/usr/bin/env bats
|
|
# Seam: frank-large — the Dankclaw iGPU endpoint on :7770.
|
|
#
|
|
# This file asserts the vLLM ROOT-UNIT shape of the 2026-10-01 cutover: the
|
|
# BF16 MiMo-V2.6-Distill-Qwen-9B payload (the same arch frank-small serves
|
|
# INT8 on the 3090) under the digest-pinned kyuz0 TheRock gfx1151 image,
|
|
# with CROSS-MODEL MTP spec decode from the Qwen3.5-9B head. The gufo-era
|
|
# laws (built-from-source runtime, GGUF schema gates, DFlash2-on-27B) retired
|
|
# with that runtime; rollback is the tag frank-large-gufo-q4kxl-20261001 and
|
|
# scripts/build-gufo.sh stays on disk for it.
|
|
#
|
|
# What the seam pins that the flags alone do not say: the measured laws of
|
|
# the 2026-10-01 window (BF16 decode 13.1 t/s is ~94% of the LPDDR5x roof,
|
|
# so spec is the only way past the wall; MTP-2 measured +51% decode against
|
|
# DFlash-8's +33%; every spec arm pays ~55% of prefill; spec arms do NOT
|
|
# reproduce the no-spec greedy text on vLLM), the roof-free shape (KV grows
|
|
# only on the 8 full-attention layers, so 4 x 262,144 fits the GTT
|
|
# aperture), the client contract difference (vLLM emits `reasoning`, not
|
|
# llama.cpp's `reasoning_content`), and the deployer's image/payload/spec
|
|
# gates with the false-green MainPID check.
|
|
|
|
load helpers
|
|
|
|
UNIT="$REPO_ROOT/scripts/frank-large.service"
|
|
SERVE="$REPO_ROOT/scripts/serve-frank-large.sh"
|
|
ENV_FILE="$REPO_ROOT/scripts/frank-large.env"
|
|
REMOTE_DEPLOY="$REPO_ROOT/scripts/deploy-remote.sh"
|
|
FETCH_TARGET="$REPO_ROOT/scripts/download-frank-large-mimo.sh"
|
|
FETCH_DRAFT="$REPO_ROOT/scripts/download-frank-large-dflash.sh"
|
|
FETCH_MTPSRC="$REPO_ROOT/scripts/download-frank-large-qwen35.sh"
|
|
|
|
@test "frank-large unit, serve script and env exist in scripts/" {
|
|
[ -f "$UNIT" ]
|
|
[ -x "$SERVE" ]
|
|
[ -f "$ENV_FILE" ]
|
|
}
|
|
|
|
@test "frank-large is boot-persistent and self-healing" {
|
|
grep -q '^WantedBy=multi-user.target$' "$UNIT"
|
|
# Restart=always, not on-failure: vLLM's API server can drain and exit 0
|
|
# after an engine fault (the frank-primary 2026-10-01 lesson), so
|
|
# on-failure would leave the endpoint dark until a human notices.
|
|
grep -q '^Restart=always$' "$UNIT"
|
|
! grep -q '^Restart=on-failure$' "$UNIT"
|
|
grep -q '^RestartSec=' "$UNIT"
|
|
# ExecStop runs `podman stop -t 120`; TimeoutStopSec MUST exceed it or
|
|
# every stop records 'Failed with result timeout' (frank-primary law).
|
|
grep -q '^ExecStop=.*podman stop -t 120 frank-large-mimo' "$UNIT"
|
|
stop_t=$(awk -F= '/^TimeoutStopSec=/{print $2}' "$UNIT")
|
|
[ -n "$stop_t" ] && [ "$stop_t" -gt 120 ]
|
|
# StartLimitIntervalSec belongs to [Unit]; a [Service] copy is an
|
|
# 'Unknown key' warning on every load.
|
|
[ "$(grep -c '^StartLimitIntervalSec=' "$UNIT")" = "1" ]
|
|
awk '/^\[Service\]/{in_svc=1} /^StartLimitIntervalSec=/{if(in_svc) exit 1}' "$UNIT"
|
|
grep -q '^User=dank$' "$UNIT"
|
|
grep -q '^LimitMEMLOCK=infinity$' "$UNIT"
|
|
grep -q '^OOMScoreAdjust=-500$' "$UNIT"
|
|
}
|
|
|
|
@test "the unit runs the repo's own serve script with the repo's env" {
|
|
grep -q '^ExecStart=/home/dank/serverAppConfigs/scripts/serve-frank-large.sh$' "$UNIT"
|
|
grep -q 'FRANK_LARGE_ENV_FILE=/home/dank/serverAppConfigs/scripts/frank-large.env' "$UNIT"
|
|
}
|
|
|
|
@test "the serve shape is direct podman run, not the retired gufo or llama shapes" {
|
|
grep -q 'podman run' "$SERVE"
|
|
grep -q 'exec podman wait "$FRANK_LARGE_CONTAINER"' "$SERVE"
|
|
# The gufo runtime and the llama.cpp prebuilt are both retired engines;
|
|
# nothing in this trio may reach for them.
|
|
! grep -q 'gufo serve' "$SERVE"
|
|
! grep -q 'llama-server' "$SERVE"
|
|
! grep -q -- '--sessions' "$SERVE"
|
|
! grep -q -- '--model-draft' "$SERVE"
|
|
! grep -q -- '--dflash-model' "$SERVE"
|
|
}
|
|
|
|
@test "the serve script gates the digest-pinned image and the complete payload" {
|
|
grep -q 'FRANK_LARGE_IMAGE' "$SERVE"
|
|
grep -q 'RepoDigests' "$SERVE"
|
|
grep -q 'FRANK_LARGE_MODEL_SHARDS' "$SERVE"
|
|
grep -q 'config.json' "$SERVE"
|
|
grep -q 'model.safetensors.index.json' "$SERVE"
|
|
# Whole /dev/dri, never individual render nodes (Dankserver 2026-09-29:
|
|
# renderD-only mounts enumerate the GPU but cannot create a context).
|
|
grep -q -- '--device /dev/dri' "$SERVE"
|
|
! grep -q 'renderD' "$SERVE"
|
|
# The env file pins by digest, not by the floating tag.
|
|
! grep -qE '^FRANK_LARGE_IMAGE=.*:latest$' "$ENV_FILE"
|
|
grep -q '^FRANK_LARGE_IMAGE=docker.io/kyuz0/vllm-therock-gfx1151@sha256:' "$ENV_FILE"
|
|
}
|
|
|
|
@test "the qualified engine shape is in the env file" {
|
|
# 2026-10-01 window: 262144 is the model's native window and the GTT
|
|
# aperture swallows it — KV grows ONLY on the 8 full-attention layers
|
|
# (32 KiB/token bf16); the 24 GDN layers hold fixed ~25 MiB state per
|
|
# running sequence. SEQS 4 keeps the four-session contract; the pool is
|
|
# shared by the scheduler instead of carved per session like gufo.
|
|
grep -q '^FRANK_LARGE_MAXLEN=262144$' "$ENV_FILE"
|
|
grep -q '^FRANK_LARGE_SEQS=4$' "$ENV_FILE"
|
|
grep -q '^FRANK_LARGE_UTIL=0.45$' "$ENV_FILE"
|
|
grep -q '^FRANK_LARGE_MODEL_DIR=/home/dank/models/MiMo-V2.6-Distill-Qwen-9B-BF16$' "$ENV_FILE"
|
|
grep -q '^FRANK_LARGE_MODEL_SHARDS=4$' "$ENV_FILE"
|
|
}
|
|
|
|
@test "the serve script wires the engine flags from the env" {
|
|
grep -q -- '--max-model-len "$FRANK_LARGE_MAXLEN"' "$SERVE"
|
|
grep -q -- '--max-num-seqs "$FRANK_LARGE_SEQS"' "$SERVE"
|
|
grep -q -- '--gpu-memory-utilization "$FRANK_LARGE_UTIL"' "$SERVE"
|
|
grep -q -- '--dtype bfloat16' "$SERVE"
|
|
grep -q -- '--trust-remote-code' "$SERVE"
|
|
grep -q -- '--ulimit memlock=-1:-1' "$SERVE"
|
|
# vllm runs INSIDE the container: the model path handed to `serve` is
|
|
# the /models mount path, never the host path (2026-10-01: the host
|
|
# path crash-looped 51 times on "Repo id must be in the form…").
|
|
grep -q 'serve "$MODEL_CONTAINER_DIR"' "$SERVE"
|
|
grep -q 'MODEL_CONTAINER_DIR="/models/' "$SERVE"
|
|
! grep -qE '^\s+serve "\$FRANK_LARGE_MODEL_DIR"' "$SERVE"
|
|
}
|
|
|
|
@test "speculative decode is cross-model MTP BY MEASUREMENT" {
|
|
# The 2026-10-01 window on this host: base 13.1 t/s, DFlash-8 17.4,
|
|
# MTP-2 19.8 (+51%). MTP acceptance is 71%/44% per position (token rate
|
|
# 57%) because the head trained on the distill's SOURCE (Qwen/Qwen3.5-9B,
|
|
# the only place mtp.* tensors exist for this family) still reads MiMo's
|
|
# hidden states. Beyond 2 draft tokens is UNMEASURED; the DFlash drafter
|
|
# stays on disk as the alternative method.
|
|
grep -q '^FRANK_LARGE_SPEC=mtp$' "$ENV_FILE"
|
|
grep -q '^FRANK_LARGE_SPEC_DIR=/home/dank/models/Qwen3.5-9B$' "$ENV_FILE"
|
|
grep -q '^FRANK_LARGE_SPEC_TOKENS=2$' "$ENV_FILE"
|
|
# The JSON is built ONLY when SPEC is set, and the draft path handed to
|
|
# the engine is the CONTAINER mount path, never the host path.
|
|
grep -q -- '--speculative-config' "$SERVE"
|
|
grep -q 'FRANK_LARGE_SPEC_CONTAINER_DIR="/models/' "$SERVE"
|
|
# The draft gate accepts BOTH weight layouts: a single
|
|
# model.safetensors (dflash) and a sharded index (the Qwen3.5-9B MTP
|
|
# head). A gate keyed only to the single file rejects the deployed arm.
|
|
grep -q 'FRANK_LARGE_SPEC_DIR/model.safetensors.index.json' "$SERVE"
|
|
# itself carries ZERO mtp.* tensors (that absence is what makes SGLang's
|
|
# NEXTN on frank-small a 0.00-acceptance ghost).
|
|
grep -q 'ZERO mtp' "$ENV_FILE"
|
|
}
|
|
|
|
@test "tools and reasoning are wired — the client contract is served" {
|
|
grep -q '^FRANK_LARGE_REASONING_PARSER=qwen3$' "$ENV_FILE"
|
|
grep -q '^FRANK_LARGE_TOOL_PARSER=qwen3_coder$' "$ENV_FILE"
|
|
grep -q -- '--reasoning-parser "$FRANK_LARGE_REASONING_PARSER"' "$SERVE"
|
|
grep -q -- '--enable-auto-tool-choice' "$SERVE"
|
|
grep -q -- '--tool-call-parser "$FRANK_LARGE_TOOL_PARSER"' "$SERVE"
|
|
# vLLM's field name is `reasoning`, NOT llama.cpp/gufo's
|
|
# `reasoning_content` — the harness models.yml compat block must match.
|
|
grep -q 'reasoning' "$ENV_FILE"
|
|
}
|
|
|
|
@test "the serve script clears stale containers including the bench identity" {
|
|
grep -q '^FRANK_LARGE_LEGACY_CONTAINERS="vllm-mimo"$' "$ENV_FILE"
|
|
grep -q '^FRANK_LARGE_CONTAINER=frank-large-mimo$' "$ENV_FILE"
|
|
grep -q 'FRANK_LARGE_LEGACY_CONTAINERS' "$SERVE"
|
|
}
|
|
|
|
@test "frank-large holds the single fleet port under its own alias" {
|
|
grep -q '^FRANK_LARGE_PUBLISH=7770:7770$' "$ENV_FILE"
|
|
grep -q '^FRANK_LARGE_SERVED_NAME=frank-large$' "$ENV_FILE"
|
|
}
|
|
|
|
@test "no bare HIP-reserved env var is defined" {
|
|
# HIP reads bare TEMP/HOST/GPU_MAX_HEAP_SIZE; the serve script sources
|
|
# this file with set -a, so an unprefixed key leaks into the engine.
|
|
! grep -qE '^(TEMP|HOST|GPU_MAX_HEAP_SIZE)=' "$ENV_FILE"
|
|
! grep -qE '^(export )?(TEMP|HOST|GPU_MAX_HEAP_SIZE)=' "$SERVE"
|
|
}
|
|
|
|
@test "the fetch scripts carry the verified payload and draft manifests" {
|
|
grep -q 'XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B' "$FETCH_TARGET"
|
|
[ "$(grep -c 'safetensors.*[0-9a-f]\{64\}' "$FETCH_TARGET")" = "4" ]
|
|
grep -q -- '--verify-only' "$FETCH_TARGET"
|
|
grep -q 'z-lab/Qwen3.5-9B-DFlash' "$FETCH_DRAFT"
|
|
grep -q -- '--verify-only' "$FETCH_DRAFT"
|
|
# The env file's payload/draft dirs must be where the fetch scripts write.
|
|
grep -q 'MiMo-V2.6-Distill-Qwen-9B-BF16' "$FETCH_TARGET"
|
|
grep -q 'Qwen3.5-9B-DFlash' "$FETCH_DRAFT"
|
|
grep -q 'Qwen/Qwen3.5-9B' "$FETCH_MTPSRC"
|
|
grep -q '/home/dank/models/Qwen3.5-9B' "$FETCH_MTPSRC"
|
|
}
|
|
|
|
@test "the gufo deploy gates are retired with the gufo runtime" {
|
|
! grep -q 'LLM_LARGE_RUNTIME' "$REMOTE_DEPLOY"
|
|
! grep -q 'LLM_LARGE_MODEL' "$REMOTE_DEPLOY"
|
|
! grep -q 'LLM_LARGE_DRAFT' "$REMOTE_DEPLOY"
|
|
! grep -q "pkill -f 'Qwen3.8-27B-DFlash2-Q4_K_M'" "$REMOTE_DEPLOY"
|
|
# build-gufo.sh stays on disk as the rollback path to the tagged engine.
|
|
[ -f "$REPO_ROOT/scripts/build-gufo.sh" ]
|
|
}
|
|
|
|
@test "the deployer gates image payload and spec then health-checks with MainPID" {
|
|
grep -q 'LLM_LARGE_CONTAINER=frank-large-mimo' "$REMOTE_DEPLOY"
|
|
grep -q 'LLM_LARGE_LEGACY_CONTAINERS="vllm-mimo"' "$REMOTE_DEPLOY"
|
|
grep -q 'FRANK_LARGE_IMAGE' "$REMOTE_DEPLOY"
|
|
grep -q "podman images --format '{{.RepoDigests}}'" "$REMOTE_DEPLOY"
|
|
grep -q 'FRANK_LARGE_MODEL_SHARDS' "$REMOTE_DEPLOY"
|
|
# The deployer's draft gate must accept the SHARDED MTP layout too —
|
|
# the 2026-10-01 first deploy died on exactly this single-file check.
|
|
grep -q 'spec_dir/model.safetensors.index.json' "$REMOTE_DEPLOY"
|
|
grep -q 'FRANK_LARGE_SPEC=' "$REMOTE_DEPLOY"
|
|
grep -q 'wait_port_free 7770' "$REMOTE_DEPLOY"
|
|
grep -q 'MainPID --value frank-large.service' "$REMOTE_DEPLOY"
|
|
}
|