launcher/recipes/glm53-flash-spark-4x/recipe.toml

# GLM-5.3-Flash NVFP4-Spark — 4x DGX Spark, the lab's current Spark quant on
# the LATEST jovian-judgement (r38 cut, 2026-09-19).
#
# Operator direction (2026-09-19): after the NAS disk death the interim
# serving mode is 2x + 4x on local NVMe with the current lab quants. This
# lane uses the latest lab engine (r38 = dev/jovian-judgement tip 8e1f1e5,
# see profiles/r38.env) instead of the legacy r12 rebuild
# (the r12 image is gone from the nodes; the r12 composition is a "phantom"
# — the lab mainline keeps the split GDN naming the current quants carry,
# and the r38 tip adds the B12X MXFP8 linear kernel that the r35-era
# flashinfer-only path lacked for the small GDN GEMMs).
#
# Model: local-inference-lab/GLM-5.3-Flash-NVFP4-Spark (154.1 GiB, snapshot
# a6082410, 09-17) — the "DSpark speculative embedded" checkpoint: the MTP
# draft head lives in the model-mtp-*.safetensors shards (no separate
# drafter download). NVFP4 routed experts + NVFP4 MTP expert layer. Split
# GDN naming (q/k/g_a/g_b/f_a/f_b/b_proj) — loads on the JJ fork lineage
# (r12..r38); the Aiden fork's fused in_proj_qkvbfg naming KeyErrors on it.
#
# Image: glm53-jovian-r38:local — the r27-lineage clone-at-pin build at the
# r38 pins (normal in-build git, no pre-staged tarballs). Flags mirror the
# proven r27-lineage GLM recipe (glm53-flash-r27-4x-dcp4, 160/176 bench)
# minus the DCP4 pair:
#   * --attention-backend B12X — required: the sm12x auto-priority list never
#     offers the B12X sparse-MLA backend for GLM's head-512
#   * no --kv-cache-dtype — let the fork pick its default (explicit fp8 /
#     fp8_ds_mla both failed auto-selection on this build lineage)
#   * --load-format instanttensor — baked into the r27-lineage images
#   * no --moe-backend — the b12x MoE routing is the fork default for this
#     quant (the FINAL BISECT "+moe b12x" gain is in the default path)
#   * speculative-config carries attention_backend B12X — the draft model's
#     vllm_config takes its attention backend from the SPECULATIVE config
#   * --language-model-only, no --enforce-eager (r27-lineage: graphs+MTP is
#     clean on the r17+ refs; the day-0 graph bug does not apply)
#
# PREREQUISITES: weights on each node's LOCAL hub (the NAS disk died
# 2026-09-19): /home/kilork/hf-local, ~/.cache/huggingface symlink
# re-pointed per node (snapshot a6082410).

engine = "vllm"
name = "GLM-5.3-Flash NVFP4-Spark (vLLM, 4x Spark, MTP-3, r38 tip)"
docker_image = "jovian-judgement:r35-glm53"  # r35 loads the Spark checkpoint (r38 pool-deadlock, see bug report)
docker_image_dir = "jovian-judgement"
build_profile = "r35"
entrypoint = ["vllm", "serve"]
model = "local-inference-lab/GLM-5.3-Flash-NVFP4-Spark"
parent_model = "zai-org/GLM-5.3-Flash"
required_nodes = 4
tensor_parallel_size = 4
pipeline_parallel_size = 1
distributed_backend = "mp"
shm_size = "32gb"

args = [
    # Stable served name — quant-agnostic (shared across glm53 quants)
    "--served-model-name", "glm53",
    "--trust-remote-code",
    "--distributed-executor-backend", "mp",
    # GMU is ONLY a startup guardrail here (KV pool is pinned below):
    # 0.88 = 107.0 GiB, below any realistic free-memory dip on GB10's
    # 121.63 GiB (the 0.90 admission flake, #48140).
    "--gpu-memory-utilization", "0.88",
    # 16 GiB KV (4x NVFP4 has the headroom: 38.5 GiB weights/node at the
    # Spark quant vs ~108 GiB usable)
    "--kv-cache-memory-bytes", "17179869184",
    "--block-size", "256",
    "--max-model-len", "524288",
    "--max-num-seqs", "16",
    "--max-num-batched-tokens", "4096",
    "--enable-chunked-prefill",
    "--async-scheduling",
    "--enable-prefix-caching",
    # B12X sparse-MLA (required for head-512 on sm_12x)
    "--attention-backend", "B12X",
    # pure I/O loader (baked into the r27-lineage images; ~5.5 min for the
    # full checkpoint off shared NFS, ~1 min off local NVMe)
    "--load-format", "instanttensor",
    # Their proven boot: no flashinfer autotune, no CuTe-DSL warmup
    "--kernel-config", """
'{"enable_cutedsl_warmup":false,"enable_flashinfer_autotune":false}'""",
    # Built-in MTP draft, 3 tokens (the r12-qualified "mtp:3" mode). The
    # draft model's vllm_config takes its attention backend from the
    # SPECULATIVE config, not the global flag — B12X mirrors the main model.
    "--speculative-config", """
'{"method":"mtp","num_speculative_tokens":3,"attention_backend":"B12X"}'""",
    "--language-model-only",
    "--tool-call-parser", "glm47",
    "--enable-auto-tool-choice",
    "--reasoning-parser", "glm45",
]

[env]
HF_HUB_OFFLINE = "1"
# RoCEnante (b12x one-shot RoCE all-reduce/all-gather) — default on for the
# r27 lanes since the 2026-09-08 A/B (4x: decode +31..+90%, quality 160 vs
# 155). Needs the RoCE fabric; single-node serves skip it silently.
VLLM_ENABLE_ROCE_ALLREDUCE = "1"
# Route mxfp8 linears away from flashinfer cutlass (rejects small GDN GEMMs: min n/k=128)
VLLM_DISABLED_KERNELS = "FlashInferCutlassMxfp8LinearKernel"
# FlashInfer JIT compiles at first boot with N parallel nvcc jobs ->
# OOM-killed on GB10 while weights+KV are resident; cap the ninja jobs.
MAX_JOBS = "2"
# NoPE-MLA rope-pad runtime gate (mod baked into the image lineage)
VLLM_MLA_NOPE_PAD_ROPE = "1"
VLLM_ENGINE_READY_TIMEOUT_S = "3600"
# Upstream launch-script env (memory + compile knobs)
PYTORCH_CUDA_ALLOC_CONF = "expandable_segments:True"
VLLM_USE_AOT_COMPILE = "0"
VLLM_USE_STANDALONE_COMPILE = "1"
VLLM_USE_FLASHINFER_SAMPLER = "1"
VLLM_USE_BREAKABLE_CUDAGRAPH = "0"
TORCH_USE_RTLD_GLOBAL = "1"
VLLM_NO_USAGE_STATS = "1"

# Architecture (aarch64 / sm_121a DGX Spark)
TORCH_CUDA_ARCH_LIST = "12.1a"
FLASHINFER_CUDA_ARCH_LIST = "12.1f"
CUTE_DSL_ARCH = "sm_121a"
CMAKE_CUDA_ARCHITECTURES = "121"
OMP_NUM_THREADS = "2"

# JIT caches persist per-node across serves (cold /root/.cache per container
# = ~20 min flashinfer/torch recompile every serve)
[[extra_mounts]]
host = "~/sm121-cache/glm53-jit-121f/b12x"
container = "/root/.cache/b12x"

[[extra_mounts]]
host = "~/sm121-cache/jit-121f/flashinfer"
container = "/root/.cache/flashinfer"

[[extra_mounts]]
host = "~/sm121-cache/jit-121f/torchinductor"
container = "/root/.cache/torchinductor"

[[extra_mounts]]
host = "~/sm121-cache/jit-121f/vllm"
container = "/root/.cache/vllm"