Add vllm/tess-dual-nvfp4: first vLLM Tess slug (fastest Tess on 2x24GB) (#679)

migtissera NVFP4 via Marlin W4A16 fallback: 62.4 tok/s decode spec-off TP=2 @131K (BENCHMARKS 2026-07-11) vs the llama.cpp entry's 57.9 with MTP — first vLLM-servable Tess on consumer cards. Ships experimental: the 8-pack (the vLLM quality program A0 baseline) + stress/soak are the promotion gate; pin-fallback rule to huginnfork NVFP4A16 documented in compose + weights manual_note. Catalog plumbing: nvfp4 weights variant; qwen35-dense added to vllm-stable supported_model_families (live-validated); froggeric template coverage extended to the tess compose (shared vendored copy, patches.yml); kv_calc_supported=false on the model (hybrid — KV on 16/64 layers, naive math ~4x over; wire at promotion). Also finishes the 2026-07-11 hybrid correction in the two spots today's sweep missed (model YAML comment, llamacpp registry status_note). Counts 63→64 / 64→65. Full suite 77/78 (submit-bench = worktree fixture artifact); diagnose-profile GREEN. Claude-Session: https://claude.ai/code/session_01EfF565T9eSLaqGzidyJ1Pm Co-authored-by: noonghunna <10742901+noonghunna@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
2026-07-12 01:00:25 +05:00
parent 36295beeae
commit 0b970ab0a7
7 changed files with 219 additions and 10 deletions
--- a/models/tess-4-27b/vllm/cache/.gitignore
+++ b/models/tess-4-27b/vllm/cache/.gitignore
@@ -0,0 +1,5 @@
+# Per-variant compile/Triton artifacts. These get regenerated on first boot
+# of any variant whose config changed; not source. Don't commit.
+*
+!.gitignore
+!README.md
--- a/models/tess-4-27b/vllm/compose/dual/nvfp4/fp8.yml
+++ b/models/tess-4-27b/vllm/compose/dual/nvfp4/fp8.yml
@@ -0,0 +1,169 @@
+# ===========================================================================
+# Profile (at-a-glance):
+#   Model:     Tess-4-27B NVFP4 (migtissera compressed-tensors: W4A4 recipe,
+#              FP8-dynamic attention/linear_attn + NVFP4 FFNs; ~19 GB.
+#              Executes as Marlin W4A16 weight-only on Ampere)
+#   Topology:  Dual 3090 PCIe (TP=2, no NVLink) — AMPERE-VALIDATED serve;
+#              native FP4 (sm_90+) untested. required_sm=9.0 (native),
+#              fallback_sm=7.5.
+#   Drafter:   none — DELIBERATE, forensically settled 2026-07-11: the grafted
+#              base-Qwen MTP head accepts 0% in vLLM (feature-space misaligned;
+#              works only token-fed in llama.cpp draft-mtp) and the EAGLE3
+#              head (~40% accept) is NET-NEGATIVE on this fast trunk (−27%).
+#              See learnings/tess-4-27b.md "2026-07-11" + club #662. Don't
+#              re-add without a retrained head.
+#   KV:        fp8 → e4m3 at scale=1.0 (no scale tensors in this checkpoint;
+#              runtime KV calibration force-disabled on qwen3_5 hybrids —
+#              measured quality-equivalent regime on this family)
+#   Vision:    yes (tower unquantized in-checkpoint) — UNTESTED on this serve;
+#              all validation is text-only
+#   Max ctx:   131K (validated envelope; weights ~9.5 GB/card leave headroom —
+#              raise MAX_MODEL_LEN at your own risk, unprofiled)
+#   Genesis:   N/A — Genesis is Qwen3-Next-specific (this is qwen3_5)
+#   Status:    🧪 Experimental
+#   Best for:  the FASTEST Tess on 2× 24 GB — 62.4 tok/s spec-off vs the
+#              llama.cpp catalog entry's 57.9 with MTP; first vLLM-servable
+#              Tess on consumer cards
+# ---------------------------------------------------------------------------
+# FIRST-PARTY VALIDATED ON AMPERE (2026-07-11 forensics week, reference 2×3090):
+#   - bench: **62.4 tok/s decode spec-off (CV 0.2%)** @131K + fp8 KV — beats
+#     the shipped llama.cpp Q4+MTP entry (57.9). BENCHMARKS.md Tess section.
+#   - Boot/serving path exercised heavily across the EAGLE3/MTP forensics
+#     (dozens of boots, coherence-clean), KV-scale probes, and template A/Bs.
+#   - 🧪 because the 8-PACK HAS NOT RUN on this serve (scheduled — "A0" in the
+#     vLLM quality program) and verify-stress/soak are unrun. Promotion gate:
+#     A0 quality vs the GGUF tier bar (115/118 ±5-7 noise) + stress + soak.
+#   - PIN RULE (maintainer, 2026-07-11): this compose pins migtissera's NVFP4
+#     (the artifact all first-party numbers were measured on). If A0 quality
+#     undershoots the GGUF bar beyond noise, the fallback pin is
+#     huginnfork/Tess-4-27B-NVFP4A16 (on-disk; W4A16-calibrated, measured
+#     KLD 0.0418 vs bf16, regrafted MTP slots) — mechanism: W4A4 weight
+#     rounding can cost quality even under Ampere's weight-only execution.
+#   - Template: the repo's chat_template.jinja is STOCK-BROKEN (developer-role
+#     crash; HF PR with the fix is in flight) — this compose pins the vendored
+#     froggeric v19 family template regardless, same as the qwen3.6 composes.
+#
+# ⚠️ Native-FP4 rigs (5090/Hopper/Blackwell): Ampere executes weight-only
+# W4A16 — your silicon also quantizes activations (true W4A4), reported to
+# cost real quality on this family (~2× KLD). All numbers here are the W4A16
+# bound. If you run this natively, rebench-full + 8-pack and report — the
+# native-W4A4 Tess datapoint doesn't exist yet.
+#
+# Run:
+#   cd <repo>/models/tess-4-27b/vllm/compose
+#   docker compose -f dual/nvfp4/fp8.yml up -d
+# ===========================================================================
+# Hardware metadata (parsed by scripts/preflight.sh):
+# Requires-min-vram-gb: 24
+# Engine-profile: vllm-stable
+# Requires-min-gpu-count: 2
+# Tensor-parallel: 2
+services:
+  vllm-tess-4-27b-dual-nvfp4:
+    # Pinned to a vLLM STABLE release (immutable, #407); the launchers inject
+    # the engine-profile image over this default (#254).
+    image: ${VLLM_IMAGE:-vllm/vllm-openai:v0.24.0}
+    container_name: "${ESTATE_CONTAINER:-vllm-tess-4-27b-dual-nvfp4}"
+    restart: ${CLUB3090_RESTART:-unless-stopped}
+    ports:
+      - "${BIND_HOST:-0.0.0.0}:${ESTATE_PORT:-${PORT:-8082}}:8000"
+    volumes:
+      - ${MODEL_DIR:-../../../../../../models-cache}:/root/.cache/huggingface
+      # torch.compile + Triton kernel caches — warm-start across boots.
+      - ../../../cache/torch_compile:/root/.cache/vllm/torch_compile_cache
+      - ../../../cache/triton:/root/.triton/cache
+      # froggeric v19 family chat template — SHARED with the qwen3.6-27b
+      # composes (same qwen3_5/qwen3.6 template family; single vendored copy,
+      # see patches.yml qwen-froggeric-chat-template). Required: the model
+      # repo's own template is stock-broken (developer-role crash).
+      - ../../../../../qwen3.6-27b/vllm/patches/froggeric-chat-template/chat_template.jinja:/etc/qwen-froggeric-chat-template.jinja:ro
+      # NVLink auto-detection — runs inside container at boot.
+      - ../../../../../../scripts/detect_nvlink.sh:/etc/club3090/detect_nvlink.sh:ro
+    environment:
+      - NVIDIA_VISIBLE_DEVICES=${ESTATE_GPUS:-${NVIDIA_VISIBLE_DEVICES:-all}}
+      # CUDA-level GPU mask — see #610; UUID form is renumbering-proof and is
+      # what CDI runtimes honor. Unset → absent (no masking).
+      - CUDA_VISIBLE_DEVICES
+      - HUGGING_FACE_HUB_TOKEN=${HF_TOKEN:-}
+      - HF_HUB_OFFLINE=${HF_HUB_OFFLINE:-${OFFLINE:-0}}
+      - TRANSFORMERS_OFFLINE=${TRANSFORMERS_OFFLINE:-${OFFLINE:-0}}
+      - VLLM_WORKER_MULTIPROC_METHOD=spawn
+      - NVLINK_MODE=${NVLINK_MODE:-auto}
+      - NCCL_CUMEM_ENABLE=0
+      # Safe PCIe default (mirrors the validated dual composes); rigs with
+      # working P2P set NCCL_P2P_DISABLE=0.
+      - NCCL_P2P_DISABLE=${NCCL_P2P_DISABLE:-1}
+      - VLLM_NO_USAGE_STATS=1
+      - OMP_NUM_THREADS=1
+      # FP8-GEMM lib guard for consumer Blackwell — launchers auto-inject
+      # VLLM_USE_DEEP_GEMM=0 there (disc #571); raw compose users set it.
+      - VLLM_USE_DEEP_GEMM
+      - PYTORCH_CUDA_ALLOC_CONF=${PYTORCH_CUDA_ALLOC_CONF:-expandable_segments:True}
+    shm_size: "16gb"
+    ipc: host
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: all
+              capabilities: [gpu]
+    entrypoint:
+      - bash
+      - -c
+      - |
+        # NVLink auto-detection (sets NCCL env vars, _NVLINK_ENABLED).
+        source /etc/club3090/detect_nvlink.sh
+        if [ "$${_NVLINK_ENABLED:-0}" = "1" ]; then
+          exec vllm serve "$$@"
+        else
+          exec vllm serve --disable-custom-all-reduce "$$@"
+        fi
+      - --
+    command:
+      - --model
+      - /root/.cache/huggingface/tess-4-27b-nvfp4
+      - --served-model-name
+      - tess-4-27b
+      - tess-4-27b-nvfp4
+      # NO --quantization flag — compressed-tensors auto-detects. Do NOT pass
+      # `--quantization modelopt` (that's the nvidia-export recipe; it fails
+      # config validation on this checkpoint).
+      - --dtype
+      - bfloat16
+      - --tensor-parallel-size
+      - "${TP:-2}"
+      - --pipeline-parallel-size
+      - "${PP:-1}"
+      - --max-model-len
+      - "${MAX_MODEL_LEN:-131072}"
+      - --gpu-memory-utilization
+      - "${GPU_MEMORY_UTILIZATION:-0.92}"
+      - --max-num-seqs
+      - "${MAX_NUM_SEQS:-2}"
+      - --max-num-batched-tokens
+      - "8192"
+      # fp8 (= e4m3) at scale=1.0 — no scale tensors in this checkpoint;
+      # measured quality-equivalent regime on this family.
+      - --kv-cache-dtype
+      - "${KV_CACHE_DTYPE:-fp8}"
+      - --trust-remote-code
+      # froggeric v19 pin (see volumes) — the model-dir template is broken.
+      - --chat-template
+      - /etc/qwen-froggeric-chat-template.jinja
+      - --reasoning-parser
+      - qwen3
+      - --default-chat-template-kwargs
+      - '{"enable_thinking": false}'
+      - --enable-auto-tool-choice
+      - --tool-call-parser
+      - qwen3_coder
+      - --enable-chunked-prefill
+      # NO --speculative-config: grafted MTP head = 0% accept in vLLM;
+      # EAGLE3 = net-negative. Forensically settled — see header.
+      - --override-generation-config
+      - '{"temperature":${TEMP:-${TEMPERATURE:-0.6}},"top_p":${TOP_P:-0.95},"top_k":${TOP_K:-20},"min_p":${MIN_P:-0.0},"repetition_penalty":${REPEAT_PENALTY:-1.0}}'
+      - --host
+      - 0.0.0.0
+      - --port
+      - "8000"
--- a/scripts/lib/profiles/compose_registry.py
+++ b/scripts/lib/profiles/compose_registry.py
@@ -893,7 +893,18 @@ COMPOSE_REGISTRY = {
        default_port=8020,
        kvcalc_key="SKIP",
        status="caveats",
-        status_note="Tess-4-27B (migtissera Q4_K_M GGUF, 16 GB) — Qwen3.5-based dense 27B instruct/agentic fine-tune on dual 3090 llama.cpp. Arch CONFIRMED qwen35-dense (standard GQA, 64 layers) from the GGUF header — same family as Deckard-40B. EXTERNAL MTP n=2 (separate mtp-*.gguf draft via --spec-draft-model, spec_method mtp_gguf) — first external-draft compose in the catalog. q4_0 KV, 262K ctx. Live-validated 2026-07-09 on server-cuda-b9246: decode ~52 narrative / 68 code tok/s (TTFT 233 ms), prefill ~1.3K tok/s; verify-stress 8/8 (NIAH ladder clean to 240,634 tok = 91% of 262K, ~5.9 GB free at deepest fill); soak-continuous PASS (0 err, 0/100 silent-empty, p50 66.4 tok/s, 96.3% retention). Quality (benchlocal --full): core 8-pack 115/150 (77%) think-off, 118/150 (79%) think-on — ties-to-edges the qwen3.6-27b dual-max reference (109) and LEADS the agentic packs (hermesagent 15/20 vs 9, cli-40 25/40 vs 20). CAVEAT: streaming tool-calls + thinking-ON -> finish=length (heavy reasoner blows the token budget before emitting the call); use thinking-OFF (shipped default) for tool/agent streaming. Trades ~1/2 the qwen-dual throughput for a quality tie/edge + vision-capable base + smaller footprint (~12.7+17.2 GB layer-split vs ~22 GB/card TP=2).",
+        status_note="Tess-4-27B (migtissera Q4_K_M GGUF, 16 GB) — Qwen3.5-based dense 27B instruct/agentic fine-tune on dual 3090 llama.cpp. Arch qwen35-dense (dense = non-MoE; HYBRID attention, 48 linear + 16 full — corrected 2026-07-11) — same family as Deckard-40B. EXTERNAL MTP n=2 (separate mtp-*.gguf draft via --spec-draft-model, spec_method mtp_gguf) — first external-draft compose in the catalog. q4_0 KV, 262K ctx. Live-validated 2026-07-09 on server-cuda-b9246: decode ~52 narrative / 68 code tok/s (TTFT 233 ms), prefill ~1.3K tok/s; verify-stress 8/8 (NIAH ladder clean to 240,634 tok = 91% of 262K, ~5.9 GB free at deepest fill); soak-continuous PASS (0 err, 0/100 silent-empty, p50 66.4 tok/s, 96.3% retention). Quality (benchlocal --full): core 8-pack 115/150 (77%) think-off, 118/150 (79%) think-on — ties-to-edges the qwen3.6-27b dual-max reference (109) and LEADS the agentic packs (hermesagent 15/20 vs 9, cli-40 25/40 vs 20). CAVEAT: streaming tool-calls + thinking-ON -> finish=length (heavy reasoner blows the token budget before emitting the call); use thinking-OFF (shipped default) for tool/agent streaming. Trades ~1/2 the qwen-dual throughput for a quality tie/edge + vision-capable base + smaller footprint (~12.7+17.2 GB layer-split vs ~22 GB/card TP=2).",
+    ),
+
+    "vllm/tess-dual-nvfp4": _entry(
+        model="tess-4-27b", weights_variant="nvfp4", workload="fast-chat",
+        engine="vllm-stable", drafter=None, kv_format="fp8_e4m3",
+        tp=2, max_ctx=131072, max_num_seqs=2, mem_util=0.92,
+        compose_path="models/tess-4-27b/vllm/compose/dual/nvfp4/fp8.yml",
+        default_port=8082, required_sm=9.0, fallback_sm=7.5,
+        kvcalc_key="SKIP",
+        status="experimental",
+        status_note="Tess-4-27B NVFP4 (migtissera compressed-tensors, W4A4 recipe; Marlin W4A16 weight-only on Ampere) at TP=2 @131K + fp8 KV, spec-off — THE FASTEST TESS on 2x24GB: 62.4 tok/s decode (CV 0.2%) vs the llama.cpp catalog entry's 57.9 with MTP, and the first vLLM-servable Tess on consumer cards (validated 2026-07-11, BENCHMARKS). Froggeric template pinned (repo template stock-broken). Drafter-less BY FORENSIC RESULT: grafted MTP head = 0% accept in vLLM (works only token-fed in llama.cpp), EAGLE3 ~40% = net-negative on this trunk (club #662). kvcalc SKIP: Tess is a qwen35 HYBRID (KV on 16/64 layers) and no hybrid kv-calc model exists for it yet — follow-up at promotion. 8-pack NOT yet run on this serve (the vLLM quality program A0 baseline) + stress/soak unrun = the promotion gate; pin-fallback rule to huginnfork NVFP4A16 if A0 undershoots the GGUF bar (learnings). No DEFAULTS row (llamacpp/tess-dual-mtp stays the tess recommendation until A0 lands).",
    ),

    # Ornith-1.0-9B — DeepReinforce agentic-coding RL fine-tune. Qwen3-Next DENSE-FFN
--- a/scripts/lib/profiles/engines/vllm-stable.yml
+++ b/scripts/lib/profiles/engines/vllm-stable.yml
@@ -15,6 +15,10 @@ supported_model_families:
                     # the ephemeral vllm-gemma4-unified image onto this engine (2026-07-01).
  - qwen3-next-hybrid
  - qwen3-next-moe
+  - qwen35-dense     # Tess-4-27B (qwen3_5 hybrid) — LIVE-VALIDATED on stock v0.24.0
+                     # 2026-07-11 (62.4 tok/s TP=2, forensics-week boots; slug
+                     # vllm/tess-dual-nvfp4). Family also covers Deckard-40B if a
+                     # vLLM path is ever added there.
 features:
  int8_per_token_head: true    # NATIVE for qwen3-next (uniform head dims) since v0.22.0 — that's why the
                               # flag is true. NOT correct for Gemma-4 (hetero heads): on v0.24.0 int8-PTH
--- a/scripts/lib/profiles/models/tess-4-27b.yml
+++ b/scripts/lib/profiles/models/tess-4-27b.yml
@@ -3,11 +3,16 @@ id: tess-4-27b
 display_name: Tess 4 27B
 family: qwen35-dense
 # Migtissera's Tess-4-27B — a Qwen3.5-based dense 27B instruct/agentic fine-tune.
-# Arch CONFIRMED from the GGUF header (general.architecture=qwen35): STANDARD GQA
-# attention, uniform across all 64 layers — NOT a Qwen3-Next/DeltaNet hybrid (no
-# GDN / linear-attention layers). Same qwen35-dense family as Qwen3.6-40B-Deckard,
-# 64 layers instead of 97. The base is VL-capable (an F16 mmproj ships alongside
-# the weights), but this catalog entry serves TEXT-ONLY — no mmproj is mounted.
+# Arch: qwen35 — ⚠️ CORRECTED 2026-07-11: this is a HYBRID (48 linear_attention
+# SSM-style + 16 full_attention layers, 3:1 interleave; model_type qwen3_5_text,
+# verified from safetensors config). The earlier "standard GQA uniform" claim
+# came from the GGUF header, which only describes the full-attention geometry.
+# NOTE for kv-calc: the attn fields below describe the 16 FULL-ATTENTION layers;
+# KV cache exists on those 16 only — naive 64-layer KV math overestimates ~4×
+# (why the vLLM entry ships kvcalc_key=SKIP pending real hybrid modeling).
+# Same qwen35-dense family as Qwen3.6-40B-Deckard (also hybrid), 64 layers vs 97.
+# The base is VL-capable (an F16 mmproj ships alongside the GGUF weights), but
+# the GGUF catalog entry serves TEXT-ONLY — no mmproj is mounted.
 # A vision variant (mtp-vision.yml) can be added once validated.
 hidden_size: 5120
 num_hidden_layers: 64
@@ -17,12 +22,26 @@ head_dim_attn: 256          # key_length == value_length == 256
 intermediate_size: 17408    # feed_forward_length
 attention_k_eq_v: false     # standard GQA — K and V are separate caches (same as Deckard/27B)
 max_ctx_supported: 262144
+kv_calc_supported: false    # HYBRID (KV on 16/64 layers) — not wired into
+                            # tools/kv-calc.py; naive 64-layer math overestimates
+                            # ~4x. Wire + calibrate at vLLM-slug promotion time.
 # rope_theta = 1e7. MTP: migtissera ships the nextn (MTP) predict-layer as a
 # SEPARATE GGUF (mtp-Tess-4-27B-Q4_K_M.gguf) — an EXTERNAL draft, engaged via
 # `--spec-draft-model <file> --spec-type draft-mtp` (drafter: tess-mtp-gguf).
 # The main quant carries NO embedded nextn head (contrast Deckard, embedded).
 vision_capable: false
 weights:
+  nvfp4:
+    path: tess-4-27b-nvfp4
+    local_subdir: tess-4-27b-nvfp4
+    size_gb: 19
+    format: compressed-tensors
+    status: experimental
+    hf_repo: migtissera/Tess-4-27B-NVFP4
+    engine: vllm
+    kind: main
+    verify_glob: "*.safetensors"
+    manual_note: "migtissera's llm-compressor NVFP4 (TRUE W4A4 recipe: 4-bit float gs16 weights AND activations on FFNs, FP8-dynamic attention/linear_attn; single safetensors, 2,672 tensors, 0 mtp — llm-compressor drops the head). On Ampere executes as Marlin W4A16 weight-only → 62.4 tok/s TP=2 @131K (2026-07-11, BENCHMARKS). Quant AUTO-DETECTS (do NOT pass --quantization modelopt). Repo chat_template is STOCK-BROKEN (developer-role crash; our HF fix PR in flight) → composes must pin froggeric. PIN RULE: this is the shipped vLLM pin (all first-party numbers measured on it); fallback = huginnfork/Tess-4-27B-NVFP4A16 (on-disk, W4A16-calibrated, KLD 0.0418) ONLY IF the A0 8-pack undershoots the GGUF tier bar beyond noise. Native-FP4 (sm_90+) quality unmeasured."
  migtissera-q4km:
    path: tess-4-27b-gguf/migtissera-q4km
    local_subdir: tess-4-27b-gguf/migtissera-q4km
--- a/scripts/lib/profiles/patches.yml
+++ b/scripts/lib/profiles/patches.yml
@@ -236,14 +236,15 @@ patches:
    status: verified

  - id: qwen-froggeric-chat-template
-    model: [qwen3.6-27b]
+    model: [qwen3.6-27b, tess-4-27b]
    files:
      - models/qwen3.6-27b/vllm/patches/froggeric-chat-template/chat_template.jinja
    load_bearing_when:
      - composes:
          - vllm/minimal
          - vllm/dual
-        reason: "Default Qwen3.6 chat template mis-handles tool-call XML / reasoning delimiters / streaming (the #145 silent-break class); the froggeric override fixes 7 default-template defects and is behavior-critical for every tool/agent compose. Carnice/Qwopus are intentional excludes (they ship their own carnice template)."
+          - vllm/tess-dual-nvfp4
+        reason: "Default Qwen3.6 chat template mis-handles tool-call XML / reasoning delimiters / streaming (the #145 silent-break class); the froggeric override fixes 7 default-template defects and is behavior-critical for every tool/agent compose. Tess-4-27B (same qwen3_5 template family) additionally NEEDS it because its repo template is stock-broken (developer-role crash). Carnice/Qwopus are intentional excludes (they ship their own carnice template)."
        evidence: "models/qwen3.6-27b/vllm/patches/froggeric-chat-template/PROVENANCE.md; docs/UPSTREAM.md (froggeric template re-eval #150 / v19 adopted PR #157)"
    delivery:               # DEPRECATED/READ-ONLY (test-only) — see header
      dockerfile_bake: false
--- a/scripts/tests/test-compose-registry-disk.sh
+++ b/scripts/tests/test-compose-registry-disk.sh
@@ -25,8 +25,8 @@ def check(cond, msg):
        print(f"FAIL: {msg}")
        failures.append(msg)

-check(len(COMPOSE_REGISTRY) == 63, f"registry has 63 entries (got {len(COMPOSE_REGISTRY)})")
-check(len(disk_paths) == 64, f"disk has 64 compose files (got {len(disk_paths)})")
+check(len(COMPOSE_REGISTRY) == 64, f"registry has 64 entries (got {len(COMPOSE_REGISTRY)})")
+check(len(disk_paths) == 65, f"disk has 65 compose files (got {len(disk_paths)})")
 check(registry_paths <= disk_paths, "all registry compose_path values exist on disk")
 parked_disk_only = disk_paths - registry_paths
 # Disk-only (non-registry) composes allowed: parked SGLang archives, plus the experimental