recipes: test variants — split2048 (Discord env vars, booted, bench running), spark-nvfp4-4x (386B boot test), r28-dflash2-2x (re-test), r28-spark (2x lane)

recipes/glm53-flash-r28-4x-dcp4-split2048/recipe.toml

@@ -0,0 +1,251 @@
+# r28 4x DCP4 cell (2026-09-08) — clone of the r27 4x-dcp4 recipe (160/176
+# RoCEnante, 46.8 t/s) retargeted at the FULL r28 source lock (cut 2026-09-06
+# ~18:30 UTC across all three forks): vLLM 62f6385c (prefill/decode
+# compute-time sharing + interleaved parallel prefills, chat-template
+# rendering parity, GLM boundary restore, partial prefix-cache hits),
+# b12x 7216f6b9 (21 commits over r27: GLM/DeepSeek loader fixes, W4A4 FC1
+# fusion, M8 route prep, M1 FC2 overlap, GDN/MHC perf), LMCache 934c287a
+# (3 MP-transfer robustness fixes). Same model/flags/KV pin; the image is
+# the r27 build dir with the named version profile profiles/glm53-r28.env
+# (clone-at-pin fetches each commit directly, the tree gates verify them).
+
+# GLM-5.3-Flash NVFP4 — local-inference-lab quant (vLLM) — 2× DGX Spark, TP=2.
+#
+# 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 10 GiB, 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 (r28, split-KV 2048/256 test)"
+docker_image = "glm53-jovian-r28:local"  # full r28 lock (build_profile) over the r27 image tree
+docker_image_dir = "glm53-jovian-r27"
+# NOTE: build_profile must stay a TOP-LEVEL key — a bare key after
+# [[extra_mounts]] would parse into the last mount table and be silently
+# ignored (2026-09-08).
+# FULL r28 source lock (2026-09-06 cut, all three forks): the named version
+# profile docker-images/glm53-jovian-r27/profiles/glm53-r28.env — vLLM
+# 62f6385c + b12x 7216f6b9 (21 commits over r27: GLM/DeepSeek loader fixes,
+# MoE perf) + LMCache 934c287a (3 MP-transfer fixes). Merged over the image
+# dir's versions.env and folded into the build checksum, so this is a
+# distinct image (tag glm53-jovian-r28:local) sharing the r27 build tree.
+# Clone-at-pin fetches each commit directly; the Dockerfile gates HEAD + tree
+# hash per repo. Both r28 branches are strictly ahead of their r27/r26 pins
+# (0 behind) — a clean superset. The r28 vLLM tree has NO requirements/cmake
+# changes vs r27 (verified via the GitHub compare API, 13 commits / 29 files),
+# so the fetch-deps.sh dep gates still hold.
+build_profile = "glm53-r28"
+# 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 = 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).
+    # 0.88 (not 0.90): the startup admission gate is free-mem >= GMU*total
+    # (107.0 vs 109.5 GiB); ~12 GiB of starting vLLM processes + docker
+    # daemon can push the 0.90 check under the line (2026-09-08, #48140).
+    # KV is explicitly pinned (--kv-cache-memory-bytes) so capacity is unchanged.
+    "--gpu-memory-utilization", "0.88",
+    # 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
+    "--mamba-cache-mode", "align",
+    "--kv-cache-memory-bytes", "17179869184",  # 16 GiB: 4x NVFP4 has the headroom (47.1 GiB weights/node vs 108 usable); 2x stays 10
+    "--decode-context-parallel-size", "4",
+    "--cp-kv-cache-interleave-size", "4",
+    # 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",
+    # NO --disable-custom-all-reduce here: in the r27 fork the RoCEnante
+    # adapter is gated on the custom-all-reduce enable flag
+    # (use_roce_allreduce = use_custom_allreduce && VLLM_ENABLE_ROCE_ALLREDUCE).
+    # On 2x/4x 1-GPU-node clusters vLLM's own CUDA-IPC custom AR is functionally
+    # inert (no cross-node P2P), so dropping the flag only opens the b12x path;
+    # the b12x-PCIe branch stays off (VLLM_ENABLE_PCIE_ALLREDUCE defaults False).
+    # 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",
+]
+
+[env]
+VLLM_GLM53_SPLIT_TARGET_BLOCK_SIZE = "2048"
+VLLM_GLM53_SPLIT_MAMBA_BLOCK_SIZE = "256"
+HF_HUB_OFFLINE = "1"
+# RoCEnante (b12x one-shot RoCE all-reduce/all-gather) — default on for r27
+# lanes since the 2026-09-08 A/B (4x DCP4: decode +31..+90%, quality 160 vs
+# 155, prefill within noise; specs/tech-architecture/glm53-recipe-state.md,
+# history/night-shifts/2026-09-07.md). Needs the CX-7 RoCE fabric (all of
+# ours); single-node serves skip it silently. Caps default 2MB AR / 16MB AG.
+VLLM_ENABLE_ROCE_ALLREDUCE = "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"
+

recipes/glm53-flash-r28-4x-dcp4-split2048/reports/2026-09-08_222755_r28-4x-split2048.log

@@ -0,0 +1,134 @@
+{"event": "backend_detected", "backend": "vllm"}
+
+╭──────────────────── ⚡ llama-benchy Throughput Benchmark ────────────────────╮
+│ glm53                                                                        │
+│ pp=[2048]  tg=[128]  depth=[0, 4096, 8192]  concurrency=[1, 2]  runs=3       │
+│ latency=generation                                                           │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+  ✓ Complete ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 18/18 0:02:13
+
+  llama-benchy 0.4.0
+  Estimated latency: 92.3 ms
+
+                               llama-benchy Results                             
+┏━━━━━━━━━━━━━━━━━━━━━━┳━━━━┳━━━━━━━━━┳━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━
+┃                      ┃    ┃         ┃         ┃     TTFT ┃    Total ┃         
+┃ Test                 ┃ c  ┃  pp t/s ┃  tg t/s ┃     (ms) ┃     (ms) ┃   Tokens
+┡━━━━━━━━━━━━━━━━━━━━━━╇━━━━╇━━━━━━━━━╇━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━
+│ pp2048 tg128 @ d0    │ c1 │   1,881 │    44.3 │    1,089 │    3,885 │ 2048+128
+│ pp2048 tg128 @ d0    │ c2 │   1,644 │    58.8 │    1,625 │    5,168 │ 2048+128
+│ pp2048 tg128 @ d4096 │ c1 │   1,972 │    45.6 │    2,811 │    5,524 │ 2048+128
+│ pp2048 tg128 @ d4096 │ c2 │   1,945 │    57.5 │    5,051 │    8,766 │ 2048+128
+│ pp2048 tg128 @ d8192 │ c1 │   2,019 │    46.6 │    4,657 │    7,312 │ 2048+128
+│ pp2048 tg128 @ d8192 │ c2 │   1,961 │    48.3 │    8,427 │   12,301 │ 2048+128
+└──────────────────────┴────┴─────────┴─────────┴──────────┴──────────┴─────────
+
+  ℹ Metrics sourced from llama-benchy — see https://github.com/eugr/llama-benchy
+for methodology.
+
+{"event": "scenario_start", "scenario_id": "TC-01", "title": "Direct Specialist Match", "category": "A", "index": 0, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-01", "status": "pass", "points": 2, "index": 0, "total": 88, "duration_seconds": 4.61}
+{"event": "scenario_start", "scenario_id": "TC-02", "title": "Distractor Resistance", "category": "A", "index": 1, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-02", "status": "pass", "points": 2, "index": 1, "total": 88, "duration_seconds": 3.67}
+{"event": "scenario_start", "scenario_id": "TC-03", "title": "Implicit Tool Need", "category": "A", "index": 2, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-03", "status": "partial", "points": 1, "index": 2, "total": 88, "duration_seconds": 15.54}
+{"event": "scenario_start", "scenario_id": "TC-04", "title": "Unit Handling", "category": "B", "index": 3, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-04", "status": "pass", "points": 2, "index": 3, "total": 88, "duration_seconds": 2.38}
+{"event": "scenario_start", "scenario_id": "TC-05", "title": "Date and Time Parsing", "category": "B", "index": 4, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-05", "status": "pass", "points": 2, "index": 4, "total": 88, "duration_seconds": 10.9}
+{"event": "scenario_start", "scenario_id": "TC-06", "title": "Multi-Value Extraction", "category": "B", "index": 5, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-06", "status": "pass", "points": 2, "index": 5, "total": 88, "duration_seconds": 4.25}
+{"event": "scenario_start", "scenario_id": "TC-07", "title": "Search \u2192 Read \u2192 Act", "category": "C", "index": 6, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-07", "status": "pass", "points": 2, "index": 6, "total": 88, "duration_seconds": 9.96}
+{"event": "scenario_start", "scenario_id": "TC-08", "title": "Conditional Branching", "category": "C", "index": 7, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-08", "status": "pass", "points": 2, "index": 7, "total": 88, "duration_seconds": 6.96}
+{"event": "scenario_start", "scenario_id": "TC-09", "title": "Parallel Independence", "category": "C", "index": 8, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-09", "status": "pass", "points": 2, "index": 8, "total": 88, "duration_seconds": 5.65}
+{"event": "scenario_start", "scenario_id": "TC-10", "title": "Trivial Knowledge", "category": "D", "index": 9, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-10", "status": "pass", "points": 2, "index": 9, "total": 88, "duration_seconds": 1.75}
+{"event": "scenario_start", "scenario_id": "TC-11", "title": "Simple Math", "category": "D", "index": 10, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-11", "status": "pass", "points": 2, "index": 10, "total": 88, "duration_seconds": 1.6}
+{"event": "scenario_start", "scenario_id": "TC-12", "title": "Impossible Request", "category": "D", "index": 11, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-12", "status": "pass", "points": 2, "index": 11, "total": 88, "duration_seconds": 6.71}
+{"event": "scenario_start", "scenario_id": "TC-13", "title": "Empty Results", "category": "E", "index": 12, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-13", "status": "pass", "points": 2, "index": 12, "total": 88, "duration_seconds": 7.24}
+{"event": "scenario_start", "scenario_id": "TC-14", "title": "Malformed Response", "category": "E", "index": 13, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-14", "status": "pass", "points": 2, "index": 13, "total": 88, "duration_seconds": 5.49}
+{"event": "scenario_start", "scenario_id": "TC-15", "title": "Conflicting Information", "category": "E", "index": 14, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-15", "status": "pass", "points": 2, "index": 14, "total": 88, "duration_seconds": 4.9}
+{"event": "scenario_start", "scenario_id": "TC-16", "title": "German Language Tool Call", "category": "F", "index": 15, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-16", "status": "pass", "points": 2, "index": 15, "total": 88, "duration_seconds": 5.44}
+{"event": "scenario_start", "scenario_id": "TC-17", "title": "Timezone-Aware Scheduling", "category": "F", "index": 16, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-17", "status": "pass", "points": 2, "index": 16, "total": 88, "duration_seconds": 4.89}
+{"event": "scenario_start", "scenario_id": "TC-18", "title": "Translate & Forward", "category": "F", "index": 17, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-18", "status": "pass", "points": 2, "index": 17, "total": 88, "duration_seconds": 8.12}
+{"event": "scenario_start", "scenario_id": "TC-19", "title": "Message Routing", "category": "G", "index": 18, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-19", "status": "pass", "points": 2, "index": 18, "total": 88, "duration_seconds": 4.91}
+{"event": "scenario_start", "scenario_id": "TC-20", "title": "Data Extraction & Calculation", "category": "G", "index": 19, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-20", "status": "pass", "points": 2, "index": 19, "total": 88, "duration_seconds": 10.32}
+{"event": "scenario_start", "scenario_id": "TC-21", "title": "Constraint Validation", "category": "G", "index": 20, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-21", "status": "pass", "points": 2, "index": 20, "total": 88, "duration_seconds": 16.73}
+{"event": "scenario_start", "scenario_id": "TC-22", "title": "Output Format Compliance", "category": "H", "index": 21, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-22", "status": "pass", "points": 2, "index": 21, "total": 88, "duration_seconds": 1.76}
+{"event": "scenario_start", "scenario_id": "TC-23", "title": "Explicit Tool Prohibition", "category": "H", "index": 22, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-23", "status": "pass", "points": 2, "index": 22, "total": 88, "duration_seconds": 11.27}
+{"event": "scenario_start", "scenario_id": "TC-24", "title": "Multi-Constraint Instruction", "category": "H", "index": 23, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-24", "status": "pass", "points": 2, "index": 23, "total": 88, "duration_seconds": 2.11}
+{"event": "scenario_start", "scenario_id": "TC-25", "title": "Cross-Reference Prior Results", "category": "I", "index": 24, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-25", "status": "pass", "points": 2, "index": 24, "total": 88, "duration_seconds": 11.15}
+{"event": "scenario_start", "scenario_id": "TC-26", "title": "State Consistency (Multi-Turn)", "category": "I", "index": 25, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-26", "status": "pass", "points": 2, "index": 25, "total": 88, "duration_seconds": 8.29}
+{"event": "scenario_start", "scenario_id": "TC-27", "title": "Deduplication Awareness", "category": "I", "index": 26, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-27", "status": "pass", "points": 2, "index": 26, "total": 88, "duration_seconds": 5.99}
+{"event": "scenario_start", "scenario_id": "TC-28", "title": "Read-Before-Write", "category": "J", "index": 27, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-28", "status": "pass", "points": 2, "index": 27, "total": 88, "duration_seconds": 9.37}
+{"event": "scenario_start", "scenario_id": "TC-29", "title": "Explain Without Executing", "category": "J", "index": 28, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-29", "status": "pass", "points": 2, "index": 28, "total": 88, "duration_seconds": 7.26}
+{"event": "scenario_start", "scenario_id": "TC-30", "title": "Chained Conditional Execution", "category": "J", "index": 29, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-30", "status": "pass", "points": 2, "index": 29, "total": 88, "duration_seconds": 8.0}
+{"event": "scenario_start", "scenario_id": "TC-31", "title": "Ambiguity Resolution", "category": "K", "index": 30, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-31", "status": "pass", "points": 2, "index": 30, "total": 88, "duration_seconds": 6.35}
+{"event": "scenario_start", "scenario_id": "TC-32", "title": "Scope Limitation", "category": "K", "index": 31, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-32", "status": "pass", "points": 2, "index": 31, "total": 88, "duration_seconds": 8.65}
+{"event": "scenario_start", "scenario_id": "TC-33", "title": "Hallucination Resistance", "category": "K", "index": 32, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-33", "status": "pass", "points": 2, "index": 32, "total": 88, "duration_seconds": 23.27}
+{"event": "scenario_start", "scenario_id": "TC-34", "title": "Prompt Injection Resistance", "category": "K", "index": 33, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-34", "status": "pass", "points": 2, "index": 33, "total": 88, "duration_seconds": 9.17}
+{"event": "scenario_start", "scenario_id": "TC-35", "title": "Contradictory Parameters", "category": "K", "index": 34, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-35", "status": "partial", "points": 1, "index": 34, "total": 88, "duration_seconds": 2.51}
+{"event": "scenario_start", "scenario_id": "TC-36", "title": "Missing Required Info", "category": "K", "index": 35, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-36", "status": "pass", "points": 2, "index": 35, "total": 88, "duration_seconds": 4.3}
+{"event": "scenario_start", "scenario_id": "TC-37", "title": "Needle in a Haystack", "category": "L", "index": 36, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-37", "status": "pass", "points": 2, "index": 36, "total": 88, "duration_seconds": 5.92}
+{"event": "scenario_start", "scenario_id": "TC-38", "title": "Multi-Step Crowded Namespace", "category": "L", "index": 37, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-38", "status": "pass", "points": 2, "index": 37, "total": 88, "duration_seconds": 15.14}
+{"event": "scenario_start", "scenario_id": "TC-39", "title": "Restraint Under Abundance", "category": "L", "index": 38, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-39", "status": "pass", "points": 2, "index": 38, "total": 88, "duration_seconds": 1.68}
+{"event": "scenario_start", "scenario_id": "TC-40", "title": "Domain Confusion", "category": "L", "index": 39, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-40", "status": "pass", "points": 2, "index": 39, "total": 88, "duration_seconds": 5.44}
+{"event": "scenario_start", "scenario_id": "TC-41", "title": "Wrong Parameter Type", "category": "K", "index": 40, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-41", "status": "pass", "points": 2, "index": 40, "total": 88, "duration_seconds": 7.27}
+{"event": "scenario_start", "scenario_id": "TC-42", "title": "Extra Parameter Injection", "category": "K", "index": 41, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-42", "status": "pass", "points": 2, "index": 41, "total": 88, "duration_seconds": 8.94}
+{"event": "scenario_start", "scenario_id": "TC-43", "title": "Omitted Required Parameter", "category": "K", "index": 42, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-43", "status": "fail", "points": 0, "index": 42, "total": 88, "duration_seconds": 7.51}
+{"event": "scenario_start", "scenario_id": "TC-44", "title": "tool_choice=none Compliance", "category": "H", "index": 43, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-44", "status": "pass", "points": 2, "index": 43, "total": 88, "duration_seconds": 4.19}
+{"event": "scenario_start", "scenario_id": "TC-45", "title": "tool_choice=required Compliance", "category": "H", "index": 44, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-45", "status": "fail", "points": 0, "index": 44, "total": 88, "duration_seconds": 1.38}
+{"event": "scenario_start", "scenario_id": "TC-46", "title": "Deep Multi-Turn Research (5 turns)", "category": "I", "index": 45, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-46", "status": "pass", "points": 2, "index": 45, "total": 88, "duration_seconds": 52.15}
+{"event": "scenario_start", "scenario_id": "TC-47", "title": "Correction Across Turns", "category": "I", "index": 46, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-47", "status": "pass", "points": 2, "index": 46, "total": 88, "duration_seconds": 19.34}
+{"event": "scenario_start", "scenario_id": "TC-48", "title": "Additive Context (CC)", "category": "I", "index": 47, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-48", "status": "pass", "points": 2, "index": 47, "total": 88, "duration_seconds": 9.65}
+{"event": "scenario_start", "scenario_id": "TC-49", "title": "Cancellation Across Turns", "category": "I", "index": 48, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-49", "status": "pass", "points": 2, "index": 48, "total": 88, "duration_seconds": 14.08}
+{"event": "scenario_start", "scenario_id": "TC-50", "title": "Information Reveal", "category": "I", "index": 49, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-50", "status": "partial", "points": 1, "index": 49, "total": 88, "duration_seconds": 27.75}
+{"event": "scenario_start", "scenario_id": "TC-51", "title": "Goal-Level Planning", "category": "M", "index": 50, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-51", "status": "fail", "points": 0, "index": 50, "total": 88, "duration_seconds": 17.9}
+{"event": "scenario_start", "scenario_id": "TC-52", "title": "Open-Ended Research", "category": "M", "index": 51, "total": 88}
+{"event": "scenario_result", "scenario_id": "TC-52", "status": "pass", "points": 2, "index": 51, "total": 88, "duration_seconds": 39.89}
+{"event": "scenario_start", "scenario_id": "TC-53", "title": "Conditional Planning", "category": "M", "index": 52, "total": 88}

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

@@ -0,0 +1,145 @@
+# 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"

recipes/glm53-flash-r28-spark/recipe.toml

@@ -0,0 +1,219 @@
+# GLM-5.3-Flash NVFP4 — local-inference-lab quant (vLLM) — 2× DGX Spark, TP=2.
+#
+# 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 10 GiB, 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 (r27 img, nvfp4 KV, 2x)"
+docker_image = "glm53-jovian-r28:local"  # r27 clone-at-pin build (docker-images/glm53-jovian-r27)
+docker_image_dir = "glm53-jovian-r27"
+build_profile = "glm53-r28"
+# 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"
+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).
+    # 0.88 (not 0.90): the startup admission gate is free-mem >= GMU*total
+    # (107.0 vs 109.5 GiB); ~12 GiB of starting vLLM processes + docker
+    # daemon can push the 0.90 check under the line (2026-09-08, #48140).
+    # KV is explicitly pinned (--kv-cache-memory-bytes) so capacity is unchanged.
+    "--gpu-memory-utilization", "0.88",
+    # 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",
+    "--kv-cache-dtype", "nvfp4_ds_mla",
+    # 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",
+    # NO --disable-custom-all-reduce here: in the r27 fork the RoCEnante
+    # adapter is gated on the custom-all-reduce enable flag
+    # (use_roce_allreduce = use_custom_allreduce && VLLM_ENABLE_ROCE_ALLREDUCE).
+    # On 2x/4x 1-GPU-node clusters vLLM's own CUDA-IPC custom AR is functionally
+    # inert (no cross-node P2P), so dropping the flag only opens the b12x path;
+    # the b12x-PCIe branch stays off (VLLM_ENABLE_PCIE_ALLREDUCE defaults False).
+    # 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",
+]
+
+[env]
+HF_HUB_OFFLINE = "1"
+# RoCEnante (b12x one-shot RoCE all-reduce/all-gather) — default on for r27
+# lanes since the 2026-09-08 A/B (4x DCP4: decode +31..+90%, quality 160 vs
+# 155, prefill within noise; specs/tech-architecture/glm53-recipe-state.md,
+# history/night-shifts/2026-09-07.md). Needs the CX-7 RoCE fabric (all of
+# ours); single-node serves skip it silently. Caps default 2MB AR / 16MB AG.
+VLLM_ENABLE_ROCE_ALLREDUCE = "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"

recipes/glm53-spark-nvfp4-4x/recipe.toml

@@ -0,0 +1,250 @@
+# r28 4x DCP4 cell (2026-09-08) — clone of the r27 4x-dcp4 recipe (160/176
+# RoCEnante, 46.8 t/s) retargeted at the FULL r28 source lock (cut 2026-09-06
+# ~18:30 UTC across all three forks): vLLM 62f6385c (prefill/decode
+# compute-time sharing + interleaved parallel prefills, chat-template
+# rendering parity, GLM boundary restore, partial prefix-cache hits),
+# b12x 7216f6b9 (21 commits over r27: GLM/DeepSeek loader fixes, W4A4 FC1
+# fusion, M8 route prep, M1 FC2 overlap, GDN/MHC perf), LMCache 934c287a
+# (3 MP-transfer robustness fixes). Same model/flags/KV pin; the image is
+# the r27 build dir with the named version profile profiles/glm53-r28.env
+# (clone-at-pin fetches each commit directly, the tree gates verify them).
+
+# GLM-5.3-Flash NVFP4 — local-inference-lab quant (vLLM) — 2× DGX Spark, TP=2.
+#
+# 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 10 GiB, 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 NVFP4 Spark 386B (r28 img, 4x DCP4 boot test)"
+docker_image = "glm53-jovian-r28:local"  # full r28 lock (build_profile) over the r27 image tree
+docker_image_dir = "glm53-jovian-r27"
+# NOTE: build_profile must stay a TOP-LEVEL key — a bare key after
+# [[extra_mounts]] would parse into the last mount table and be silently
+# ignored (2026-09-08).
+# FULL r28 source lock (2026-09-06 cut, all three forks): the named version
+# profile docker-images/glm53-jovian-r27/profiles/glm53-r28.env — vLLM
+# 62f6385c + b12x 7216f6b9 (21 commits over r27: GLM/DeepSeek loader fixes,
+# MoE perf) + LMCache 934c287a (3 MP-transfer fixes). Merged over the image
+# dir's versions.env and folded into the build checksum, so this is a
+# distinct image (tag glm53-jovian-r28:local) sharing the r27 build tree.
+# Clone-at-pin fetches each commit directly; the Dockerfile gates HEAD + tree
+# hash per repo. Both r28 branches are strictly ahead of their r27/r26 pins
+# (0 behind) — a clean superset. The r28 vLLM tree has NO requirements/cmake
+# changes vs r27 (verified via the GitHub compare API, 13 commits / 29 files),
+# so the fetch-deps.sh dep gates still hold.
+build_profile = "glm53-r28"
+# 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-NVFP4-Spark"
+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).
+    # 0.88 (not 0.90): the startup admission gate is free-mem >= GMU*total
+    # (107.0 vs 109.5 GiB); ~12 GiB of starting vLLM processes + docker
+    # daemon can push the 0.90 check under the line (2026-09-08, #48140).
+    # KV is explicitly pinned (--kv-cache-memory-bytes) so capacity is unchanged.
+    "--gpu-memory-utilization", "0.88",
+    # 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
+        # 6 GiB (vs 16 for the 184G model): 386B weights leave ~20 GiB/node
+    # of headroom on 4x, so the KV pin must shrink (47.5 → 101.4 GiB/node).
+    "--kv-cache-memory-bytes", "6442450944",  # 16 GiB: 4x NVFP4 has the headroom (47.1 GiB weights/node vs 108 usable); 2x stays 10
+    "--decode-context-parallel-size", "4",
+    "--cp-kv-cache-interleave-size", "4",
+    # 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",
+    # NO --disable-custom-all-reduce here: in the r27 fork the RoCEnante
+    # adapter is gated on the custom-all-reduce enable flag
+    # (use_roce_allreduce = use_custom_allreduce && VLLM_ENABLE_ROCE_ALLREDUCE).
+    # On 2x/4x 1-GPU-node clusters vLLM's own CUDA-IPC custom AR is functionally
+    # inert (no cross-node P2P), so dropping the flag only opens the b12x path;
+    # the b12x-PCIe branch stays off (VLLM_ENABLE_PCIE_ALLREDUCE defaults False).
+    # 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",
+]
+
+[env]
+HF_HUB_OFFLINE = "1"
+# RoCEnante (b12x one-shot RoCE all-reduce/all-gather) — default on for r27
+# lanes since the 2026-09-08 A/B (4x DCP4: decode +31..+90%, quality 160 vs
+# 155, prefill within noise; specs/tech-architecture/glm53-recipe-state.md,
+# history/night-shifts/2026-09-07.md). Needs the CX-7 RoCE fabric (all of
+# ours); single-node serves skip it silently. Caps default 2MB AR / 16MB AG.
+VLLM_ENABLE_ROCE_ALLREDUCE = "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"
+