noonghunnaandClaude Fable 5
e8bcfd8da1
launch.sh/switch.sh now detect GPU arch and export KV_CACHE_DTYPE= fp8_e4m3 (native FP8 compute) on sm_89+ cards for the pilot slugs (vllm/dual, vllm/minimal). The injected value is the hardware profiles' dormant kv_format_default.balanced -- one source of truth shared with the pull gates; the Ampere no-op is data equality (3090-class balanced = fp8_e5m2 = compose default -> nothing emitted), not a code branch. Injection guards (all load-bearing): - pilot allowlist only; expansion gated on the #246 cross-rig A/B - per-variant: registry kv_format == fp8_e5m2 only (int8-PTH/TQ/bf16 slugs never touched -- compressed-tensors weights reject fp8 KV) - vLLM-family variants only; explicit user KV_CACHE_DTYPE= wins; unmapped cards / heterogeneous rigs / no nvidia-smi -> no injection - direct `docker compose up` keeps Ampere-safe compose defaults Rides the existing resolve-variant-pin export seam (new optional --gpu-spec); VLLM_ATTENTION_BACKEND is whitelisted but ships no value (vLLM auto-detect stays the default until measured). Preflight banner names the detected arch class. Consistency fixes the injection exposed: - gates.py _ARCH_KERNEL_SM: fp8_e4m3 9.0 -> 8.9 (vLLM's real floor is SM89+; we'd otherwise inject e4m3 on 4090s our own pull-gate calls unloadable) + new nvfp4: 10.0 row (v0.24.0 literal had NO gate -- a 3090 pull of an nvfp4 config wouldn't have been rejected) - Blackwell hardware profiles: nvfp4 declared as CANDIDATE capability (engine list unchanged until validated -- gates take the intersection) - kv-calc: projected nvfp4 rows (bytes/elem + activation coefs), calibration unchanged Docs: HARDWARE.md new section, KV_MATH/QUANTIZATION/DTYPE_MATRIX rows (incl. retiring the Genesis-era "e4m3 undertuned per #51" advisory in favor of the A/B). Validated: 8-case injection matrix in test-launch-compat; live no-op on the real 2x3090 (spec built, nothing injected); faked 4090 through the real bash detection path emits e4m3; compose interpolation both ways; kv-calc --calibration green; full scripts gate 64/64. Co-Authored-By: Claude Fable 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01EfF565T9eSLaqGzidyJ1Pm
147 lines
6.1 KiB
Bash
Executable File
147 lines
6.1 KiB
Bash
Executable File
#!/usr/bin/env bash
|
|
set -euo pipefail
|
|
|
|
ROOT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/../.." && pwd)"
|
|
HELPER="${ROOT_DIR}/scripts/lib/profiles/launch_compat.py"
|
|
GPU_3090='0|RTX_3090|24576|8.6'
|
|
MTP_SHA="01d4d1ad375dc5854779c593eee093bcebb0cada"
|
|
CLEAN_SHA="bf610c2f56764e1b30bc6065f4ceace3d6e59036"
|
|
DFLASH_SHA="e47c98ef7a38792996e452ef53914e21e41928e9"
|
|
|
|
assert_contains() {
|
|
local haystack="$1"
|
|
local needle="$2"
|
|
if [[ "$haystack" != *"$needle"* ]]; then
|
|
echo "ASSERTION FAILED: expected output to contain: $needle" >&2
|
|
echo "--- output ---" >&2
|
|
echo "$haystack" >&2
|
|
exit 1
|
|
fi
|
|
}
|
|
|
|
assert_not_contains() {
|
|
local haystack="$1"
|
|
local needle="$2"
|
|
if [[ "$haystack" == *"$needle"* ]]; then
|
|
echo "ASSERTION FAILED: expected output not to contain: $needle" >&2
|
|
echo "--- output ---" >&2
|
|
echo "$haystack" >&2
|
|
exit 1
|
|
fi
|
|
}
|
|
|
|
out="$(python3 "$HELPER" filter-candidates \
|
|
--variants vllm/dual,vllm/minimal,llamacpp/default \
|
|
--model qwen3.6-27b \
|
|
--gpu-spec "$GPU_3090" \
|
|
--tp 1 \
|
|
--pp 1 \
|
|
--workload fast-chat)"
|
|
assert_contains "$out" "vllm/minimal"
|
|
assert_not_contains "$out" "vllm/dual"
|
|
|
|
out="$(python3 "$HELPER" filter-candidates \
|
|
--variants vllm/minimal,llamacpp/default,llamacpp/mtp \
|
|
--model qwen3.6-27b \
|
|
--gpu-spec "$GPU_3090" \
|
|
--tp 1 \
|
|
--pp 1 \
|
|
--stable)"
|
|
assert_contains "$out" "vllm/minimal"
|
|
assert_contains "$out" "llamacpp/default"
|
|
assert_contains "$out" "llamacpp/mtp"
|
|
|
|
if out="$(python3 "$HELPER" validate-variant \
|
|
--variant vllm/gemma-mtp-tp1 \
|
|
--gpu-spec "$GPU_3090" \
|
|
--tp 2 \
|
|
--pp 1 \
|
|
--no-project-vram 2>&1)"; then
|
|
echo "ASSERTION FAILED: invalid Gemma single-card profile unexpectedly passed" >&2
|
|
echo "$out" >&2
|
|
exit 1
|
|
fi
|
|
assert_contains "$out" "C1: tp=2 * pp=1 = 2 != 1 cards selected"
|
|
assert_contains "$out" "C5: kv_format=fp8_e4m3 not supported by hardware: rtx-3090"
|
|
|
|
out="$(python3 "$HELPER" validate-variant \
|
|
--variant vllm/minimal \
|
|
--gpu-spec "$GPU_3090" \
|
|
--tp 1 \
|
|
--pp 1 \
|
|
--no-project-vram \
|
|
--verbose 2>&1)"
|
|
assert_contains "$out" "Pass 1 fits()"
|
|
assert_contains "$out" "Resolved compose: vllm/minimal"
|
|
assert_contains "$out" "Pass 2 fits()"
|
|
|
|
out="$(python3 "$HELPER" resolve-engine-pin --engine-id vllm-nightly-mtp --format shell)"
|
|
assert_contains "$out" "VLLM_NIGHTLY_SHA=${MTP_SHA}"
|
|
|
|
if out="$(python3 "$HELPER" resolve-engine-pin --engine-id vllm-pip-baseline --format shell 2>&1)"; then
|
|
echo "ASSERTION FAILED: pip-only vllm-pip-baseline unexpectedly resolved as a docker nightly" >&2
|
|
echo "$out" >&2
|
|
exit 1
|
|
fi
|
|
assert_contains "$out" "install.spec is not a docker image"
|
|
|
|
out="$(python3 "$HELPER" resolve-variant-pin --variant vllm/dual --format shell)"
|
|
assert_contains "$out" "VLLM_IMAGE=vllm/vllm-openai:v0.24.0"
|
|
|
|
|
|
out="$(python3 "$HELPER" resolve-variant-pin --variant vllm/gemma-int8-mtp --format shell)"
|
|
assert_contains "$out" "VLLM_IMAGE=vllm/vllm-openai:v0.22.0"
|
|
|
|
# --- #246 arch-aware KV injection matrix (resolve-variant-pin --gpu-spec) ----
|
|
GPU_4090='0|NVIDIA GeForce RTX 4090|24564|8.9'
|
|
GPU_5090X2='0|NVIDIA GeForce RTX 5090|32607|12.0;1|NVIDIA GeForce RTX 5090|32607|12.0'
|
|
|
|
# ampere -> NOTHING injected (compose defaults; the no-op is data equality)
|
|
out="$(python3 "$HELPER" resolve-variant-pin --variant vllm/dual --format shell --gpu-spec "$GPU_3090")"
|
|
assert_not_contains "$out" "KV_CACHE_DTYPE"
|
|
# ada / blackwell pilot slugs -> native-fp8 swap
|
|
out="$(python3 "$HELPER" resolve-variant-pin --variant vllm/dual --format shell --gpu-spec "$GPU_4090")"
|
|
assert_contains "$out" "KV_CACHE_DTYPE=fp8_e4m3"
|
|
out="$(python3 "$HELPER" resolve-variant-pin --variant vllm/minimal --format shell --gpu-spec "$GPU_5090X2")"
|
|
assert_contains "$out" "KV_CACHE_DTYPE=fp8_e4m3"
|
|
# non-pilot slug (same kv_format) -> no injection until the #246 A/B expands the pilot
|
|
out="$(python3 "$HELPER" resolve-variant-pin --variant vllm/qwen-27b-dual-fast --format shell --gpu-spec "$GPU_4090")"
|
|
assert_not_contains "$out" "KV_CACHE_DTYPE"
|
|
# quant-specific KV slug -> never overridden (compressed-tensors reject fp8 KV)
|
|
out="$(python3 "$HELPER" resolve-variant-pin --variant vllm/gemma-int8-mtp --format shell --gpu-spec "$GPU_4090")"
|
|
assert_not_contains "$out" "KV_CACHE_DTYPE"
|
|
# explicit user env pin wins
|
|
out="$(KV_CACHE_DTYPE=fp8_e5m2 python3 "$HELPER" resolve-variant-pin --variant vllm/dual --format shell --gpu-spec "$GPU_4090")"
|
|
assert_not_contains "$out" "KV_CACHE_DTYPE"
|
|
# heterogeneous rig -> no single right answer -> no injection
|
|
out="$(python3 "$HELPER" resolve-variant-pin --variant vllm/dual --format shell --gpu-spec "${GPU_3090};1|NVIDIA GeForce RTX 4090|24564|8.9")"
|
|
assert_not_contains "$out" "KV_CACHE_DTYPE"
|
|
# unmapped card -> degrade to compose defaults, never an error
|
|
out="$(python3 "$HELPER" resolve-variant-pin --variant vllm/dual --format shell --gpu-spec "0|Weird GPU|8192|7.0")"
|
|
assert_contains "$out" "VLLM_IMAGE=vllm/vllm-openai:v0.24.0"
|
|
assert_not_contains "$out" "KV_CACHE_DTYPE"
|
|
echo " ok: #246 arch-aware KV injection matrix (8 cases)"
|
|
|
|
if command -v docker >/dev/null 2>&1 && docker compose version >/dev/null 2>&1; then
|
|
out="$(VLLM_NIGHTLY_SHA="$CLEAN_SHA" docker compose -f "$ROOT_DIR/models/qwen3.6-27b/vllm/compose/dual/autoround-int4/fp8-mtp.yml" config 2>/dev/null)"
|
|
assert_contains "$out" "image: vllm/vllm-openai:v0.24.0"
|
|
|
|
out="$(VLLM_NIGHTLY_SHA="$CLEAN_SHA" VLLM_IMAGE=vllm/vllm-openai:latest docker compose -f "$ROOT_DIR/models/qwen3.6-27b/vllm/compose/dual/autoround-int4/fp8-mtp.yml" config 2>/dev/null)"
|
|
assert_contains "$out" "image: vllm/vllm-openai:latest"
|
|
fi
|
|
|
|
out="$(python3 - <<'PY'
|
|
from scripts.lib.profiles.compat import InstanceSpec
|
|
from scripts.lib.profiles.estate_cli import compose_env
|
|
|
|
clean = compose_env(InstanceSpec(name="qwen", compose_name="vllm/dual", gpu_indices=(0, 1), port=8010))
|
|
gemma = compose_env(InstanceSpec(name="gemma", compose_name="vllm/gemma-int8-mtp", gpu_indices=(0, 1), port=8032))
|
|
print(clean["VLLM_IMAGE"])
|
|
print(gemma["VLLM_IMAGE"])
|
|
PY
|
|
)"
|
|
assert_contains "$out" "vllm/vllm-openai:v0.24.0" # clean (vllm/dual → vllm-stable) bumped to v0.24.0
|
|
assert_contains "$out" "vllm/vllm-openai:v0.22.0" # gemma (vllm/gemma-int8-mtp → vllm-gemma-stable) stays v0.22.0
|
|
|
|
echo "test-launch-compat: ok"
|