ODS V3 Pre-Release: Public testing and refinement ahead of the official V3 launch. Turn your PC, Mac, or Linux box into a private AI server.
Python ai-agentsamdcomfyuidockerllama-cppllmlocal-ain8nnvidiaopen-webuiragself-hostedspeech-to-textstrix-halotext-to-speechworkflow-automation
#!/bin/bash
# ============================================================================
# Test: resolve_tier_config() — tier-map.sh
# ============================================================================
# Sources the actual tier-map.sh and verifies each tier resolves to the
# correct LLM_MODEL, GGUF_FILE, and MAX_CONTEXT.
#
# Run: bash tests/test-tier-map.sh
# ============================================================================
set -uo pipefail
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
PASS=0
FAIL=0
# Minimal stubs for dependencies
error() { echo "ERROR: $*" >&2; return 1; }
# Source the module under test
source "$SCRIPT_DIR/installers/lib/tier-map.sh"
source "$SCRIPT_DIR/lib/safe-env.sh"
load_selector_env() {
load_model_selector_env_from_output <<< "$1"
}
assert_eq() {
local label="$1" expected="$2" actual="$3"
if [[ "$expected" == "$actual" ]]; then
echo " PASS: $label"
((PASS++))
else
echo " FAIL: $label (expected '$expected', got '$actual')"
((FAIL++))
fi
}
run_tier() {
local tier_val="$1"
TIER="$tier_val"
# Reset globals
TIER_NAME="" LLM_MODEL="" GGUF_FILE="" GGUF_URL="" MAX_CONTEXT=""
GPU_BACKEND="" N_GPU_LAYERS=""
resolve_tier_config
}
echo "=== Testing resolve_tier_config() ==="
echo ""
# --- Tier 0: Lightweight ---
echo "Tier 0 (Lightweight):"
run_tier 0
assert_eq "TIER_NAME" "Lightweight" "$TIER_NAME"
assert_eq "LLM_MODEL" "qwen3.5-2b" "$LLM_MODEL"
assert_eq "GGUF_FILE" "Qwen3.5-2B-Q4_K_M.gguf" "$GGUF_FILE"
assert_eq "MAX_CONTEXT" "8192" "$MAX_CONTEXT"
echo ""
# --- Tier 1: Entry Level ---
echo "Tier 1 (Entry Level):"
run_tier 1
assert_eq "TIER_NAME" "Entry Level" "$TIER_NAME"
assert_eq "MODEL_PROFILE_EFFECTIVE" "qwen" "$MODEL_PROFILE_EFFECTIVE"
assert_eq "LLM_MODEL" "qwen3.5-9b" "$LLM_MODEL"
assert_eq "GGUF_FILE" "Qwen3.5-9B-Q4_K_M.gguf" "$GGUF_FILE"
assert_eq "MAX_CONTEXT" "16384" "$MAX_CONTEXT"
echo ""
# --- Tier 2: Prosumer ---
echo "Tier 2 (Prosumer):"
run_tier 2
assert_eq "TIER_NAME" "Prosumer" "$TIER_NAME"
assert_eq "MODEL_PROFILE_EFFECTIVE" "qwen" "$MODEL_PROFILE_EFFECTIVE"
assert_eq "LLM_MODEL" "qwen3.5-9b" "$LLM_MODEL"
assert_eq "GGUF_FILE" "Qwen3.5-9B-Q4_K_M.gguf" "$GGUF_FILE"
assert_eq "MAX_CONTEXT" "32768" "$MAX_CONTEXT"
echo ""
# --- Tier 3: Pro ---
echo "Tier 3 (Pro):"
run_tier 3
assert_eq "TIER_NAME" "Pro" "$TIER_NAME"
assert_eq "MODEL_PROFILE_EFFECTIVE" "qwen" "$MODEL_PROFILE_EFFECTIVE"
assert_eq "LLM_MODEL" "qwen3.5-27b" "$LLM_MODEL"
assert_eq "GGUF_FILE" "Qwen3.5-27B-Q4_K_M.gguf" "$GGUF_FILE"
assert_eq "MAX_CONTEXT" "65536" "$MAX_CONTEXT"
assert_eq "LLM_MODEL_SIZE_MB" "16700" "$LLM_MODEL_SIZE_MB"
echo ""
# --- Tier 4: Enterprise ---
echo "Tier 4 (Enterprise):"
run_tier 4
assert_eq "TIER_NAME" "Enterprise" "$TIER_NAME"
assert_eq "MODEL_PROFILE_EFFECTIVE" "qwen" "$MODEL_PROFILE_EFFECTIVE"
assert_eq "LLM_MODEL" "qwen3.6-35b-a3b" "$LLM_MODEL"
assert_eq "GGUF_FILE" "Qwen3.6-35B-A3B-UD-Q4_K_M.gguf" "$GGUF_FILE"
assert_eq "MAX_CONTEXT" "131072" "$MAX_CONTEXT"
assert_eq "LLM_MODEL_SIZE_MB" "21110" "$LLM_MODEL_SIZE_MB"
echo ""
# --- NV_ULTRA ---
echo "NV_ULTRA (NVIDIA Ultra 90GB+):"
unset HOST_ARCH
run_tier NV_ULTRA
assert_eq "TIER_NAME" "NVIDIA Ultra (90GB+)" "$TIER_NAME"
assert_eq "MODEL_PROFILE_EFFECTIVE" "qwen" "$MODEL_PROFILE_EFFECTIVE"
assert_eq "LLM_MODEL" "qwen3-coder-next" "$LLM_MODEL"
assert_eq "GGUF_FILE" "qwen3-coder-next-Q4_K_M.gguf" "$GGUF_FILE"
assert_eq "MAX_CONTEXT" "131072" "$MAX_CONTEXT"
echo ""
echo "NV_ULTRA (aarch64 A3B substitution):"
HOST_ARCH=arm64
run_tier NV_ULTRA
assert_eq "TIER_NAME" "NVIDIA Ultra (90GB+, aarch64 — A3B substitution)" "$TIER_NAME"
assert_eq "MODEL_PROFILE_EFFECTIVE" "qwen" "$MODEL_PROFILE_EFFECTIVE"
assert_eq "LLM_MODEL" "qwen3.6-35b-a3b" "$LLM_MODEL"
assert_eq "GGUF_FILE" "Qwen3.6-35B-A3B-UD-Q4_K_M.gguf" "$GGUF_FILE"
assert_eq "MAX_CONTEXT" "131072" "$MAX_CONTEXT"
unset HOST_ARCH
echo ""
# --- SH_LARGE ---
# SH_LARGE substitutes to 35B-A3B for the same unified-memory reason as
# NV_ULTRA aarch64 — see tier-map.sh SH_LARGE block.
echo "SH_LARGE (Strix Halo 90+):"
run_tier SH_LARGE
assert_eq "TIER_NAME" "Strix Halo 90+" "$TIER_NAME"
assert_eq "MODEL_PROFILE_EFFECTIVE" "qwen" "$MODEL_PROFILE_EFFECTIVE"
assert_eq "LLM_MODEL" "qwen3.6-35b-a3b" "$LLM_MODEL"
assert_eq "GGUF_FILE" "Qwen3.6-35B-A3B-UD-Q4_K_M.gguf" "$GGUF_FILE"
assert_eq "MAX_CONTEXT" "131072" "$MAX_CONTEXT"
echo ""
# --- SH_COMPACT ---
echo "SH_COMPACT (Strix Halo Compact):"
run_tier SH_COMPACT
assert_eq "TIER_NAME" "Strix Halo Compact" "$TIER_NAME"
assert_eq "MODEL_PROFILE_EFFECTIVE" "qwen" "$MODEL_PROFILE_EFFECTIVE"
assert_eq "LLM_MODEL" "qwen3.6-35b-a3b" "$LLM_MODEL"
assert_eq "GGUF_FILE" "Qwen3.6-35B-A3B-UD-Q4_K_M.gguf" "$GGUF_FILE"
assert_eq "MAX_CONTEXT" "131072" "$MAX_CONTEXT"
echo ""
# --- ARC ---
echo "ARC (Intel Arc ≥12 GB, e.g. A770 16 GB):"
run_tier ARC
assert_eq "TIER_NAME" "Intel Arc" "$TIER_NAME"
assert_eq "MODEL_PROFILE_EFFECTIVE" "qwen" "$MODEL_PROFILE_EFFECTIVE"
assert_eq "LLM_MODEL" "qwen3.5-9b" "$LLM_MODEL"
assert_eq "GGUF_FILE" "Qwen3.5-9B-Q4_K_M.gguf" "$GGUF_FILE"
assert_eq "MAX_CONTEXT" "32768" "$MAX_CONTEXT"
assert_eq "GPU_BACKEND" "sycl" "$GPU_BACKEND"
assert_eq "N_GPU_LAYERS" "99" "$N_GPU_LAYERS"
echo ""
# --- ARC_LITE ---
echo "ARC_LITE (Intel Arc <12 GB, e.g. A750 8 GB / A380 6 GB):"
run_tier ARC_LITE
assert_eq "TIER_NAME" "Intel Arc Lite" "$TIER_NAME"
assert_eq "MODEL_PROFILE_EFFECTIVE" "qwen" "$MODEL_PROFILE_EFFECTIVE"
assert_eq "LLM_MODEL" "qwen3.5-4b" "$LLM_MODEL"
assert_eq "GGUF_FILE" "Qwen3.5-4B-Q4_K_M.gguf" "$GGUF_FILE"
assert_eq "MAX_CONTEXT" "16384" "$MAX_CONTEXT"
assert_eq "GPU_BACKEND" "sycl" "$GPU_BACKEND"
assert_eq "N_GPU_LAYERS" "99" "$N_GPU_LAYERS"
echo ""
# --- Invalid tier should fail ---
echo "Invalid tier (should fail):"
if TIER="INVALID" resolve_tier_config 2>/dev/null; then
echo " FAIL: Invalid tier did not return error"
((FAIL++))
else
echo " PASS: Invalid tier returned error"
((PASS++))
fi
echo ""
# --- GGUF_URL should be set for all tiers ---
echo "GGUF_URL populated for all tiers:"
for t in 0 1 2 3 4 NV_ULTRA SH_LARGE SH_COMPACT ARC ARC_LITE; do
run_tier "$t"
if [[ -n "$GGUF_URL" && "$GGUF_URL" == https://* ]]; then
echo " PASS: Tier $t has valid GGUF_URL"
((PASS++))
else
echo " FAIL: Tier $t missing or invalid GGUF_URL"
((FAIL++))
fi
done
echo ""
# --- Gemma 4 profile opt-in ---
echo "Gemma 4 profile (tier 2):"
MODEL_PROFILE=gemma4 run_tier 2
assert_eq "MODEL_PROFILE_EFFECTIVE" "gemma4" "$MODEL_PROFILE_EFFECTIVE"
assert_eq "LLM_MODEL" "gemma-4-e4b-it" "$LLM_MODEL"
assert_eq "GGUF_FILE" "gemma-4-E4B-it-Q4_K_M.gguf" "$GGUF_FILE"
assert_eq "LLAMA_SERVER_IMAGE" "ghcr.io/ggml-org/llama.cpp:server-cuda-b11429@sha256:883915ea20a4b350e98f7284098f4d400c21e9b366e248d397cd42d90602a84c" "$LLAMA_SERVER_IMAGE"
assert_eq "LLAMA_CPP_RELEASE_TAG_OVERRIDE" "b9014" "$LLAMA_CPP_RELEASE_TAG_OVERRIDE"
unset MODEL_PROFILE
echo ""
echo "Auto profile (tier 0 fallback):"
MODEL_PROFILE=auto run_tier 0
assert_eq "MODEL_PROFILE_EFFECTIVE" "qwen" "$MODEL_PROFILE_EFFECTIVE"
assert_eq "LLM_MODEL" "qwen3.5-2b" "$LLM_MODEL"
unset MODEL_PROFILE
echo ""
echo "Auto profile (tier 3 prefers Gemma with safer context):"
MODEL_PROFILE=auto run_tier 3
assert_eq "MODEL_PROFILE_EFFECTIVE" "gemma4" "$MODEL_PROFILE_EFFECTIVE"
assert_eq "LLM_MODEL" "gemma-4-26b-a4b-it" "$LLM_MODEL"
assert_eq "MAX_CONTEXT" "16384" "$MAX_CONTEXT"
unset MODEL_PROFILE
echo ""
echo "Auto profile (tier 4 prefers Gemma with safer context):"
MODEL_PROFILE=auto run_tier 4
assert_eq "MODEL_PROFILE_EFFECTIVE" "gemma4" "$MODEL_PROFILE_EFFECTIVE"
assert_eq "LLM_MODEL" "gemma-4-31b-it" "$LLM_MODEL"
assert_eq "MAX_CONTEXT" "65536" "$MAX_CONTEXT"
unset MODEL_PROFILE
echo ""
echo "Auto profile (SH_COMPACT prefers Gemma with safer context):"
MODEL_PROFILE=auto run_tier SH_COMPACT
assert_eq "MODEL_PROFILE_EFFECTIVE" "gemma4" "$MODEL_PROFILE_EFFECTIVE"
assert_eq "LLM_MODEL" "gemma-4-26b-a4b-it" "$LLM_MODEL"
assert_eq "MAX_CONTEXT" "65536" "$MAX_CONTEXT"
unset MODEL_PROFILE
echo ""
echo "Explicit qwen profile (tier 2):"
MODEL_PROFILE=qwen run_tier 2
assert_eq "MODEL_PROFILE_EFFECTIVE" "qwen" "$MODEL_PROFILE_EFFECTIVE"
assert_eq "LLM_MODEL" "qwen3.5-9b" "$LLM_MODEL"
assert_eq "GGUF_FILE" "Qwen3.5-9B-Q4_K_M.gguf" "$GGUF_FILE"
unset MODEL_PROFILE
echo ""
echo "Catalog selector (amd64 NV_ULTRA keeps coder-next):"
_selector_env="$(python3 "$SCRIPT_DIR/scripts/select-model.py" \
--catalog "$SCRIPT_DIR/config/model-library.json" \
--backend nvidia \
--memory-type discrete \
--vram-mb 98304 \
--ram-gb 128 \
--profile qwen \
--tier NV_ULTRA \
--host-arch amd64 \
--installable-only \
--env)"
LLM_MODEL="" GGUF_FILE="" MODEL_RECOMMENDATION_POLICY=""
load_selector_env "$_selector_env"
assert_eq "SELECTOR_LLM_MODEL" "qwen3-coder-next" "$LLM_MODEL"
assert_eq "SELECTOR_GGUF_FILE" "qwen3-coder-next-Q4_K_M.gguf" "$GGUF_FILE"
assert_eq "SELECTOR_POLICY" "context-aware-curated-fit-v2" "$MODEL_RECOMMENDATION_POLICY"
echo ""
# Qwen 3.6 27B is a fleet A/B candidate (install_recommendation=false). The
# curated priority and the installable filter keep the 24-32GB default
# unchanged. Cover both the tier-3 size ceiling
# (non-Pixel hosts) and the uncapped Pixel-default path.
echo "Catalog selector (24-32GB discrete GPUs keep Qwen 3.5 27B; Qwen 3.6 27B stays a candidate):"
for _vram_mb in 24576 32607; do
for _backend in nvidia amd; do
for _max_size_mb in 18600 0; do
_selector_env="$(python3 "$SCRIPT_DIR/scripts/select-model.py" \
--catalog "$SCRIPT_DIR/config/model-library.json" \
--backend "$_backend" \
--memory-type discrete \
--vram-mb "$_vram_mb" \
--ram-gb 64 \
--profile qwen \
--tier 3 \
--max-size-mb "$_max_size_mb" \
--host-arch amd64 \
--installable-only \
--env)"
LLM_MODEL="" GGUF_FILE="" MAX_CONTEXT="" MODEL_RECOMMENDED_ALTERNATIVES=""
load_selector_env "$_selector_env"
_case="${_backend} ${_vram_mb}MB max-size ${_max_size_mb}"
assert_eq "SELECTOR_LLM_MODEL ($_case)" "qwen3.5-27b" "$LLM_MODEL"
assert_eq "SELECTOR_GGUF_FILE ($_case)" "Qwen3.5-27B-Q4_K_M.gguf" "$GGUF_FILE"
# The catalog's operating default is 64K: the 27B is selected at
# the context the Hermes floor used to raise it to (tower1/tower3).
assert_eq "SELECTOR_CONTEXT ($_case)" "65536" "$MAX_CONTEXT"
case ";$MODEL_RECOMMENDED_ALTERNATIVES" in
*";qwen3.6-27b-ud-q4-k-xl:"*)
echo " FAIL: SELECTOR_ALTERNATIVES ($_case) offers the Qwen 3.6 27B candidate"
((FAIL++)) ;;
*)
echo " PASS: SELECTOR_ALTERNATIVES ($_case) omits the Qwen 3.6 27B candidate"
((PASS++)) ;;
esac
done
done
done
unset _vram_mb _backend _max_size_mb _case
echo ""
echo "Catalog selector (arm64 NV_ULTRA preserves A3B substitution):"
_selector_env="$(python3 "$SCRIPT_DIR/scripts/select-model.py" \
--catalog "$SCRIPT_DIR/config/model-library.json" \
--backend nvidia \
--memory-type unified \
--vram-mb 0 \
--ram-gb 128 \
--profile qwen \
--tier NV_ULTRA \
--host-arch arm64 \
--installable-only \
--env)"
LLM_MODEL="" GGUF_FILE="" MODEL_RECOMMENDATION_POLICY=""
load_selector_env "$_selector_env"
assert_eq "SELECTOR_LLM_MODEL" "qwen3.6-35b-a3b" "$LLM_MODEL"
assert_eq "SELECTOR_GGUF_FILE" "Qwen3.6-35B-A3B-UD-Q4_K_M.gguf" "$GGUF_FILE"
assert_eq "SELECTOR_POLICY" "context-aware-curated-fit-v2+spark-aarch64-nv-ultra-a3b-v1" "$MODEL_RECOMMENDATION_POLICY"
echo ""
# Strix Halo (SH_LARGE, AMD Ryzen AI MAX+ 395, 124 GB unified) hit the same
# coder-next pathology as Spark aarch64. The selector should substitute to
# 35B-A3B via the unified-memory policy (distinct tag from the Spark one).
echo "Catalog selector (amd unified SH_LARGE substitutes coder-next → A3B):"
_selector_env="$(python3 "$SCRIPT_DIR/scripts/select-model.py" \
--catalog "$SCRIPT_DIR/config/model-library.json" \
--backend amd \
--memory-type unified \
--vram-mb 0 \
--ram-gb 124 \
--profile qwen \
--tier SH_LARGE \
--host-arch amd64 \
--installable-only \
--env)"
LLM_MODEL="" GGUF_FILE="" MODEL_RECOMMENDATION_POLICY=""
load_selector_env "$_selector_env"
assert_eq "SELECTOR_LLM_MODEL" "qwen3.6-35b-a3b" "$LLM_MODEL"
assert_eq "SELECTOR_GGUF_FILE" "Qwen3.6-35B-A3B-UD-Q4_K_M.gguf" "$GGUF_FILE"
assert_eq "SELECTOR_POLICY" "context-aware-curated-fit-v2+unified-memory-coder-next-a3b-v1" "$MODEL_RECOMMENDATION_POLICY"
echo ""
echo "Catalog selector (16GB Apple unified qwen keeps ranked fit):"
_selector_env="$(python3 "$SCRIPT_DIR/scripts/select-model.py" \
--catalog "$SCRIPT_DIR/config/model-library.json" \
--backend metal \
--memory-type unified \
--vram-mb 0 \
--ram-gb 16 \
--profile qwen \
--tier 1 \
--host-arch arm64 \
--installable-only \
--env)"
LLM_MODEL="" GGUF_FILE="" MODEL_RECOMMENDATION_SOURCE="" MODEL_RECOMMENDATION_POLICY=""
load_selector_env "$_selector_env"
assert_eq "SELECTOR_LLM_MODEL" "qwen3.5-9b" "$LLM_MODEL"
assert_eq "SELECTOR_GGUF_FILE" "Qwen3.5-9B-Q4_K_M.gguf" "$GGUF_FILE"
assert_eq "SELECTOR_SOURCE" "catalog_fit_pre_download" "$MODEL_RECOMMENDATION_SOURCE"
assert_eq "SELECTOR_POLICY" "context-aware-curated-fit-v2" "$MODEL_RECOMMENDATION_POLICY"
echo ""
echo "Catalog selector (32GB Apple unified picks the Pixel-verified 9B at 64K, not phi-4 at 16K):"
_selector_env="$(python3 "$SCRIPT_DIR/scripts/select-model.py" \
--catalog "$SCRIPT_DIR/config/model-library.json" \
--backend metal \
--memory-type unified \
--vram-mb 0 \
--ram-gb 32 \
--profile qwen \
--tier 2 \
--host-arch arm64 \
--installable-only \
--env)"
LLM_MODEL="" GGUF_FILE="" MODEL_RECOMMENDATION_SOURCE="" MODEL_RECOMMENDATION_POLICY=""
load_selector_env "$_selector_env"
assert_eq "SELECTOR_LLM_MODEL" "qwen3.5-9b" "$LLM_MODEL"
assert_eq "SELECTOR_GGUF_FILE" "Qwen3.5-9B-Q4_K_M.gguf" "$GGUF_FILE"
assert_eq "SELECTOR_SOURCE" "catalog_fit_pre_download" "$MODEL_RECOMMENDATION_SOURCE"
assert_eq "SELECTOR_POLICY" "context-aware-curated-fit-v2" "$MODEL_RECOMMENDATION_POLICY"
echo ""
echo "Catalog selector (--max-size-mb bounds tier 2 on 48GB Apple unified — issue #1881):"
_selector_env="$(python3 "$SCRIPT_DIR/scripts/select-model.py" \
--catalog "$SCRIPT_DIR/config/model-library.json" \
--backend apple \
--memory-type unified \
--vram-mb 0 \
--ram-gb 48 \
--profile qwen \
--tier 2 \
--max-size-mb 5760 \
--host-arch arm64 \
--installable-only \
--env)"
LLM_MODEL="" GGUF_FILE="" MODEL_RECOMMENDATION_SOURCE="" MODEL_RECOMMENDATION_REASON=""
load_selector_env "$_selector_env"
assert_eq "SELECTOR_LLM_MODEL" "qwen3.5-9b" "$LLM_MODEL"
assert_eq "SELECTOR_GGUF_FILE" "Qwen3.5-9B-Q4_K_M.gguf" "$GGUF_FILE"
case "$MODEL_RECOMMENDATION_REASON" in
*"Bounded by --tier 2's model size ceiling"*) echo " PASS: SELECTOR_REASON mentions ceiling"; ((PASS++)) ;;
*) echo " FAIL: SELECTOR_REASON missing ceiling note (got '$MODEL_RECOMMENDATION_REASON')"; ((FAIL++)) ;;
esac
echo ""
echo "Catalog selector (--max-size-mb absent still auto-upgrades; 48GB Apple gets the 35B-A3B MoE):"
_selector_env="$(python3 "$SCRIPT_DIR/scripts/select-model.py" \
--catalog "$SCRIPT_DIR/config/model-library.json" \
--backend apple \
--memory-type unified \
--vram-mb 0 \
--ram-gb 48 \
--profile qwen \
--tier 2 \
--host-arch arm64 \
--installable-only \
--env)"
LLM_MODEL="" GGUF_FILE=""
load_selector_env "$_selector_env"
assert_eq "SELECTOR_LLM_MODEL" "qwen3.6-35b-a3b" "$LLM_MODEL"
echo ""
echo "Catalog selector (tiny --max-size-mb falls back to smallest fitting model, not smallest overall):"
_selector_env="$(python3 "$SCRIPT_DIR/scripts/select-model.py" \
--catalog "$SCRIPT_DIR/config/model-library.json" \
--backend apple \
--memory-type unified \
--vram-mb 0 \
--ram-gb 48 \
--profile qwen \
--tier 0 \
--max-size-mb 1500 \
--host-arch arm64 \
--installable-only \
--env)"
LLM_MODEL="" GGUF_FILE=""
load_selector_env "$_selector_env"
assert_eq "SELECTOR_LLM_MODEL" "qwen3.5-2b" "$LLM_MODEL"
echo ""
echo "Catalog selector (amd unified SH_COMPACT uses ranked A3B without override):"
_selector_env="$(python3 "$SCRIPT_DIR/scripts/select-model.py" \
--catalog "$SCRIPT_DIR/config/model-library.json" \
--backend amd \
--memory-type unified \
--vram-mb 0 \
--ram-gb 64 \
--profile qwen \
--tier SH_COMPACT \
--host-arch amd64 \
--installable-only \
--env)"
LLM_MODEL="" GGUF_FILE="" MODEL_RECOMMENDATION_SOURCE="" MODEL_RECOMMENDATION_POLICY=""
load_selector_env "$_selector_env"
assert_eq "SELECTOR_LLM_MODEL" "qwen3.6-35b-a3b" "$LLM_MODEL"
assert_eq "SELECTOR_GGUF_FILE" "Qwen3.6-35B-A3B-UD-Q4_K_M.gguf" "$GGUF_FILE"
assert_eq "SELECTOR_SOURCE" "catalog_fit_pre_download" "$MODEL_RECOMMENDATION_SOURCE"
assert_eq "SELECTOR_POLICY" "context-aware-curated-fit-v2" "$MODEL_RECOMMENDATION_POLICY"
echo ""
echo "Catalog selector (8GB NVIDIA qwen uses the live-proven 64K Q8-KV interactive profile):"
_selector_env="$(python3 "$SCRIPT_DIR/scripts/select-model.py" \
--catalog "$SCRIPT_DIR/config/model-library.json" \
--backend nvidia \
--memory-type discrete \
--vram-mb 8188 \
--ram-gb 15 \
--profile qwen \
--tier 1 \
--host-arch amd64 \
--installable-only \
--env)"
LLM_MODEL="" GGUF_FILE="" MAX_CONTEXT="" MODEL_RUNTIME_PROFILE="" LLAMA_ARG_N_CPU_MOE="" LLAMA_ARG_FLASH_ATTN="" LLAMA_ARG_CACHE_TYPE_K="" LLAMA_ARG_CACHE_TYPE_V="" LLAMA_ARG_CHECKPOINT_EVERY_NT=""
load_selector_env "$_selector_env"
assert_eq "SELECTOR_LLM_MODEL" "qwen3.5-9b" "$LLM_MODEL"
assert_eq "SELECTOR_GGUF_FILE" "Qwen3.5-9B-Q4_K_M.gguf" "$GGUF_FILE"
assert_eq "SELECTOR_CONTEXT" "65536" "$MAX_CONTEXT"
assert_eq "SELECTOR_RUNTIME_PROFILE" "nvidia-8gb-64k-q8-kv" "$MODEL_RUNTIME_PROFILE"
assert_eq "SELECTOR_N_CPU_MOE" "" "$LLAMA_ARG_N_CPU_MOE"
assert_eq "SELECTOR_FLASH_ATTN" "on" "$LLAMA_ARG_FLASH_ATTN"
assert_eq "SELECTOR_CACHE_K" "q8_0" "$LLAMA_ARG_CACHE_TYPE_K"
assert_eq "SELECTOR_CACHE_V" "q8_0" "$LLAMA_ARG_CACHE_TYPE_V"
assert_eq "SELECTOR_CHECKPOINTS" "" "$LLAMA_ARG_CHECKPOINT_EVERY_NT"
echo ""
echo "Catalog selector (8GB NVIDIA gemma uses upstream catalog fit):"
_selector_env="$(python3 "$SCRIPT_DIR/scripts/select-model.py" \
--catalog "$SCRIPT_DIR/config/model-library.json" \
--backend nvidia \
--memory-type discrete \
--vram-mb 8188 \
--ram-gb 31 \
--profile gemma4 \
--tier 1 \
--host-arch amd64 \
--installable-only \
--env)"
LLM_MODEL="" GGUF_FILE="" MAX_CONTEXT="" MODEL_RUNTIME_PROFILE="" LLAMA_ARG_N_CPU_MOE="" LLAMA_ARG_CACHE_TYPE_V=""
load_selector_env "$_selector_env"
assert_eq "SELECTOR_LLM_MODEL" "gemma-4-e4b-it" "$LLM_MODEL"
assert_eq "SELECTOR_GGUF_FILE" "gemma-4-E4B-it-Q4_K_M.gguf" "$GGUF_FILE"
# Sliding-window layout: 64K costs ~6.1 GiB (1 GiB of KV), within 8 GB.
assert_eq "SELECTOR_CONTEXT" "65536" "$MAX_CONTEXT"
assert_eq "SELECTOR_RUNTIME_PROFILE" "" "$MODEL_RUNTIME_PROFILE"
assert_eq "SELECTOR_N_CPU_MOE" "" "$LLAMA_ARG_N_CPU_MOE"
assert_eq "SELECTOR_CACHE_V" "" "$LLAMA_ARG_CACHE_TYPE_V"
echo ""
echo "Catalog selector excludes Hugging Face imports from a mixed catalog:"
_selector_catalog="$(mktemp)"
trap 'rm -f "$_selector_catalog"' EXIT
python3 - "$_selector_catalog" <<'PY'
import json
import sys
models = [
{
"id": "curated-model",
"llm_model_name": "curated-model",
"family": "qwen",
"gguf_file": "curated.gguf",
"gguf_url": "https://huggingface.co/ods/curated/resolve/main/curated.gguf",
"size_mb": 1000,
"vram_required_gb": 2,
"context_length": 65536,
},
{
"id": "imported-model",
"source": "huggingface",
"llm_model_name": "imported-model",
"family": "qwen",
"gguf_file": "imported.gguf",
"gguf_url": "https://huggingface.co/community/imported/resolve/main/imported.gguf",
"size_mb": 7000,
"vram_required_gb": 7,
"context_length": 131072,
},
]
with open(sys.argv[1], "w", encoding="utf-8") as handle:
json.dump({"models": models}, handle)
PY
_selector_env="$(python3 "$SCRIPT_DIR/scripts/select-model.py" \
--catalog "$_selector_catalog" \
--backend nvidia \
--memory-type discrete \
--vram-mb 8192 \
--ram-gb 32 \
--profile qwen \
--tier 1 \
--host-arch amd64 \
--installable-only \
--env)"
rm -f "$_selector_catalog"
trap - EXIT
LLM_MODEL="" GGUF_FILE=""
load_selector_env "$_selector_env"
assert_eq "SELECTOR_CURATED_SOURCE" "curated-model" "$LLM_MODEL"
echo ""
if python3 "$SCRIPT_DIR/tests/test-gemma4-artifact-pins.py"; then
((PASS++))
else
((FAIL++))
fi
echo ""
# --- Summary ---
echo "==============================="
echo "Results: $PASS passed, $FAIL failed"
echo "==============================="
[[ $FAIL -eq 0 ]] && exit 0 || exit 1