launcher/recipes/glm53-flash-r28-dflash2-2x/recipe.toml

# DFlash2 re-test on the r28 full stack (2026-09-08, operator: "может он
# заработал"). The 2026-09-06 blocker (C4 parent-page stride assert + FP8-KV
# drafter backend rejection) targeted the r27 tree; r28's vLLM carries
# fa4720ea "fix(cache): restore GLM boundaries across DCP and DFlash groups"
# + the #646 2048/256 page split, so the geometry may now line up.
# Recipe = the r27-dflash2 2x shape (deleted at 17a91a9) retargeted:
# method "dflash", incoai/GLM-5.3-Flash-DFlash2 draft, k=7, B12X backend,
# block-size 256, 10 GiB KV pin (2x profile).
#
engine = "vllm"
name = "GLM-5.3-Flash NVFP4 (r28 img, DFlash2 k=7, 2x re-test)"
docker_image = "glm53-jovian-r28:local"  # full r28 lock via build_profile
build_profile = "glm53-r28"
docker_image_dir = "glm53-jovian-r27"
# Required: the clean r12 image follows the solar contract (ENTRYPOINT []
# + CMD bash), so the recipe carries the entrypoint (the old day-0-based
# r12 image had ["vllm","serve"] baked in, which is why this was absent
# until the clean rebuild).
entrypoint = ["vllm", "serve"]
model = "local-inference-lab/GLM-5.3-Flash-NVFP4"
required_nodes = 2
tensor_parallel_size = 2
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): 0.90
    # demands 109.46 GiB free on GB10's 121.63 GiB, but driver reserve +
    # page cache routinely leave ~108.5 -> flaky
    # "Free memory ... less than desired" boot deaths (this was the class
    # of the old "54-second death"), and post-crash reruns see ~106.6 GiB.
    # 0.82 = 99.7 GiB, below any realistic dip; does NOT change what gets
    # allocated (weights ~93 + pinned KV 10 + activations).
    "--gpu-memory-utilization", "0.90",
    # Pinned 10 GiB KV (≈1.16M tokens) — GMU sizing has no headroom at
    # 92.72 GiB weights/node; pinning also keeps warmup off the swap path
    "--kv-cache-memory-bytes", "10737418240",
    # No --kv-cache-dtype: let the r12 fork pick its default (the explicit
    # fp8 / fp8_ds_mla both failed auto-selection on this build).
    # The GLM5Next model wires the B12X sparse-MLA backend (sm_120/sm_121)
    # when the global backend is B12X (glm5next/nvidia/attention.py:
    # _select_sparse_backend); the sm12x auto-priority list never offers it
    # (FLASHINFER_MLA_SPARSE_SM120 only supports head 576, not GLM's 512).
    "--attention-backend", "B12X",
    # pure I/O loader (kept across every config; 171 GB in ~5.5 min vs ~16 NFS)
    "--load-format", "instanttensor",
    # [R17 FULL-CLUSTER TEST] lil manifest backends (quality gate: hardmode
    # must stay ≥88; r12-era this cluster dropped it to 74 — r17 may carry
    # the fix through — one boot decides).
    "--block-size", "256",
    # FINAL BISECT (2026-09-03, hardmode 176pt, r17 img):
    #   knobs-only 87: d0 c1 24.1, d8192 23.8
    #   +moe b12x  88: d0 c1 23.5, d8192 26.6  <- SHIP (2 pts d8192)
    #   +linear b12x 44: W4A4 linear kernel destroys numerics (param-precision 0%)
    #   +align     88: but d0 c1 18.6 (-5) — perf-negative, dropped
    # (all three together = 44-74; the lab needs its private packed package)
    "--max-model-len", "524288",
    "--max-num-seqs", "16",
    "--max-num-batched-tokens", "4096",
    "--enable-chunked-prefill",
    "--async-scheduling",
    "--enable-prefix-caching",
    "--disable-custom-all-reduce",
    # 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 — their
    # x86 numbers: 254 tok/s CC1, 2.37 accept; the checkpoint embeds up to
    # 5 via the model-mtp-0000x shards but 3 is the verified depth). This is
    # the checkpoint's own MTP head (Glm5NextMTPModel), NOT dspark — dspark
    # is the DeepSeek-V4 family's mechanism; GLM's other r12 mode is
    # dflash2:7 with a separate trained draft checkpoint.
    # The draft model's vllm_config takes its attention backend from the
    # SPECULATIVE config, not the global flag (eagle/utils.py:
    # _make_eagle_draft_vllm_config) — B12X here mirrors the main model's.
    # r18 used draft_model/FLASH_ATTN/kv auto — both rejected on that fork
# (FP8-KV-drafter backend check + C4-page stride assert). r27 qualifies
# dflash2 upstream (the #646 2048+256 page split is in this tree); the
# drafter/depth follow the working aiden GB10 lane (method "dflash").
"--speculative-config", """
'{"method":"dflash","model":"incoai/GLM-5.3-Flash-DFlash2","num_speculative_tokens":7}'""",
    "--language-model-only",
    # MTP + CUDA graphs degrades OUTPUT QUALITY on the glm53 day-0 dev build
    # (forum #381350/#44, CONFIRMED by A/B bench 2026-08-28: same seed/flags,
    # graphs on = 89/100, eager = 92/100; decode unchanged 28.4 t/s @ c1,
    # c2 -7%, TTFT +1.6s). The b12x glm52 stack does NOT need this (its
    # graphs+MTP combo benches clean at 100/100) — bug is build-specific.
    # graphs moved BEFORE tool-call-parser; enforce-eager removed for the
    # r17-graphs test (bisect A2 was 88/100 eager; graphs+MTP fix is public
    # in r17 refs? verify by bench).
    "--tool-call-parser", "glm47",
    "--enable-auto-tool-choice",
    "--reasoning-parser", "glm45",
]

[env]
HF_HUB_OFFLINE = "1"
# FlashInfer JIT (mxfp8_gemm_cutlass_sm120, ~12 kernels) compiles at first
# boot with 12 parallel nvcc jobs -> OOM-killed (code 137) on GB10's
# 121 GB unified while the 93 GB weights + 10 GB KV pin are resident.
# MAX_JOBS caps ninja parallelism so the JIT fits in the remaining ~20 GB.
MAX_JOBS = "2"
# NoPE-MLA rope-pad runtime gate (mod baked into the image)
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"

[[extra_mounts]]
host = "~/sm121-cache/glm53-jit-121f/b12x"
container = "/root/.cache/b12x"

# JIT caches persist per-node across serves (was: cold /root/.cache per
# container -> ~20 min flashinfer/torch recompile every serve). The host
# dirs are shared between lanes of the same arch — each cache
# self-namespaces by its engine version key internally.
[[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"