recipes: glm53-flash-nvfp4-r18-4x-nvfp4kv-spark

recipes/glm53-flash-nvfp4-r18-4x-nvfp4kv-spark/recipe.toml

@@ -0,0 +1,211 @@
+# GLM-5.3-Flash NVFP4 — local-inference-lab quant (vLLM) — 2× DGX Spark, TP=2.
+# BLOCKED 2026-09-06: the -Spark checkpoint's MXFP8 MTP-expert layer fails at
+# TP4: "ValueError: Problem size is not supported for mm_mxfp8" (the non-Spark
+# lab quant is fine at 4x — it is the Spark quant's MXFP8 tiling that does not
+# survive TP4 sharding). The Spark checkpoint's working lane is the 2x cell
+# (glm53-flash-nvfp4-r18-2x-nvfp4kv-spark: 158/176, best 2x result).
+#
+# local-inference-lab mixed quant: NVFP4 routed experts (layers 3–44,
+# group-16) + MXFP8 MTP experts (layer 45) + W4A4 input scales. Best agent
+# quant so far: tool-eval-bench 89/100 (vs 85 LibertAIDAI, 90 official FP8).
+# Model: https://huggingface.co/local-inference-lab/GLM-5.3-Flash-NVFP4
+#        (~186 GB → 92.72 GiB weights/node; KV pinned at 16 GiB (4x), see below)
+#
+# Image: glm53-flash-nvfp4-lab:local (docker-images/glm53-flash-nvfp4-lab/ —
+# day-0 arm64 image + NoPE-MLA mod + the four FujitsuPolycom lab-checkpoint
+# patches: model naming shim, modelopt MTP namespace fix, persistent_topk
+# CC-12.x guards). Their launch config is mirrored below.
+# PROVENANCE: https://github.com/FujitsuPolycom/glm53-flash-tp2-spark
+#             (Apache-2.0; see patches/UPSTREAM.md) + forum 381350 #41.
+#
+# Deviations from their launch script (intentional):
+#   * no --kda-prefill-backend flashkda (needs their base-image patch — skipped)
+#   * --kv-cache-dtype fp8_ds_mla instead of `fp8` (same packed layout; this
+#     spelling is what the NoPE-MLA mod booted with on the LibertAIDAI stack)
+#   * --language-model-only + no mm-renderer patch/env (text-only)
+#   * no bundled NCCL/LD_PRELOAD, no NCCL_* fabric block (our switched fabric
+#     is configured via [nccl.ds4] + launcher GID auto-detect)
+#   * no --moe-backend marlin: their proven stack uses the default backend for
+#     THIS quant and reports clean evals. Sanity-check output for token loops
+#     anyway (the silent FP4 MoE corruption gotcha) before trusting a new node.
+#   * flashinfer autotune disabled via --kernel-config (their proven boot);
+#     VLLM_ENGINE_READY_TIMEOUT_S stays high regardless.
+#
+# Boot success markers (engine log, in order):
+#   "Model loading took 92.72 GiB" → "[quantprobe] ... algo=MXFP8" (MTP fix
+#   active; algo=None means broken) → "GPU KV cache size: 1,164,369 tokens" →
+#   "Application startup complete". ~700 s sequential weight load.
+#
+# Test request: chat completions need "chat_template_kwargs":{"reasoning_effort"
+# :"low"} — omitted/other values select MAX reasoning effort (empty content on
+# short max_tokens).
+#
+# Optional: instanttensor loader (~30 s weight load vs ~700 s):
+#   FROM glm53-flash-nvfp4-lab:local + pip install --no-deps --no-build-isolation
+#   instanttensor==0.1.9, then add --load-format instanttensor (--no-deps is
+#   required: resolution would downgrade the image's NCCL).
+#
+# PREREQUISITES: weights cached on both nodes:
+# [A/B CELL 2026-09-03] ship recipe verbatim, checkpoint swapped to the lab
+# non-Spark NVFP4 (isolates checkpoint from image lineage vs the lab-image cell).
+# [BISECT A2 2026-09-03] EXACT 90/100 recipe (236ce5a) + instanttensor + JIT
+# mounts, on the CURRENT eugr-patched image. Isolates image lineage from the
+# execution-env/flag cluster (AOT/V2/b12x/card-env). Comparator: 124/138.
+# GLM-5.3-Flash NVFP4-Spark — local-inference-lab quant with the speculative
+# layers embedded — 2× DGX Spark, TP=2.
+#
+# local-inference-lab/GLM-5.3-Flash-NVFP4-Spark (rev 06972cdee7280cec849aedf05a02062b3e870246):
+# NVFP4 routed experts (layers 3+, group-16) + NVFP4 MTP expert layer in the
+# separate model-mtp-0000x.safetensors shards (mtp_tokens: 5 per their
+# lil.yaml), amax/nonexpert/inputscales side shards, Glm5NextForConditionalGeneration.
+# THE "DSpark speculative embedded" checkpoint — no separate DFlash2 download
+# needed; lil.yaml: default_speculator mtp, dflash2_tokens 7 (r12 doc fr: their
+# standalone-Q: MTP:3 254 tok/s CC1, 333 Sieve, 94 steps/s).
+#
+# Image: glm53-flash-nvfp4-lab:local (same proven day-0 arm64 stack + NoPE-MLA
+# mod + modelopt MTP namespace patch used by the 92/100 lab lane). The MTP
+# layer here is NVFP4 (not MXFP8) — the quantprobe marker may print
+# algo=NVFP4/- different; algo=None = patch missed (fall back to 4p67 image).
+#
+# Deviations from lab (unchanged mechanics): same fabric/env/GMU/KV pinning.
+# lil.yaml launch mirrors: block 256, disable_flashinfer_autotune, mm data
+# mode — our text-only flags identical to glm53-flash-nvfp4-lab.
+#
+# Boot markers: "Model loading took ~XX GiB" → "[quantprobe] ..." →
+# "GPU KV cache size" → "Application startup complete".
+#
+# PREREQUISITES: weights on the NAS hub (175 GiB): hf download local-inference-lab/GLM-5.3-Flash-NVFP4-Spark
+
+engine = "vllm"
+name = "GLM-5.3-Flash NVFP4-Spark (r18 img, nvfp4 KV, 4x 16G)"
+docker_image = "glm53-jovian-r18:local"  # r18 fork build (see docker-images/glm53-jovian-r18)
+docker_image_dir = "glm53-jovian-r18"
+# 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-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): 0.90
+    # 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 16 GiB KV (4x has the headroom: ~47 GiB weights/node vs 108 usable); 2x stays 10
+    "--kv-cache-memory-bytes", "17179869184",  # 16 GiB: 4x NVFP4 has the headroom (47.1 GiB weights/node vs 108 usable); 2x stays 10
+    "--kv-cache-dtype", "nvfp4_ds_mla",
+    # 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.
+    "--speculative-config", """
+'{"method":"mtp","num_speculative_tokens":3,"attention_backend":"B12X"}'""",
+    "--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",
+]
+archived = true
+
+[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"