training(eliza-1): wire qwen3.5-27b SFT launch path + extend stage_base_v1_candidate to 27b

Lands the missing pieces so the eliza-1-27b cloud SFT + HF publish chain is
addressable from train_vast.sh and stage_base_v1_candidate.py. No live run
launched here — the swarm's 0_8b path on Nebius is still working through
retries, and the only currently-listed Vast B200-1x offer is $7.5/hr ×
~50h projected wall = ~$375 (over the operator's $200 cap). Nebius H200-1x
(per the registry entry's extras={"nebius_machine": "H200-1x"}) remains the
cheaper launch target once the swarm proves the path end-to-end.

scripts/lib/vast.py:
- Add b200-1x target (1× NVIDIA B200, ≈183 GB, min_per_gpu_ram_gb=170).
  Cheapest single-GPU fit on Vast for qwen3.5-27b's 130 GB working-set
  budget at seq=32k with apollo_mini + grad-ckpt + Liger CE.

scripts/train_vast.sh:
- Wire qwen3.5-27b → b200-1x, FSDP_WORLD_SIZE=1 (was falling through to
  the catch-all blackwell6000-2x default; that worked but projected ~$458
  / 208h at the current $/hr, badly off the registry's intent).
- qwen3.6-27b legacy stays on b200-2x (still the right target for the
  larger 64k-seq context budget on that backbone).

scripts/publish/stage_base_v1_candidate.py:
- Add "27b" to --tier choices, REQUIRED_KERNELS_BY_TIER, RAM_BUDGET_MB
  (mirrors eliza1_manifest.REQUIRED_KERNELS_BY_TIER["27b"] +
  publish/orchestrator.RAM_BUDGET_BY_TIER["27b"]).
- Introduce QWEN3_PARAMS_BY_TIER lookup; replace the two hardcoded
  `'1.7B' if tier=='1_7b' else '0.6B'` ternaries with the lookup so the
  lineage/provenance/README blocks render correctly for 27b.
- Make the cuda/rocm kernel-verify caveats tier-aware: for 27b both are
  tier-supported (per SUPPORTED_BACKENDS_BY_TIER["27b"]); rocm stays
  "skipped/needs-hardware" because the build host has no AMD GPU.
- Make the drafter target-meta note tier-aware (no longer hardcoded to
  "Upstream Qwen3-0.6B GGUF used as the DFlash drafter for the 1.7B
  target").
- Document the Q8_0 voice-asset gap: VOICE_QUANT_BY_TIER["27b"]="Q8_0"
  but elizaos/eliza-1-assets only carries Q4_K_M voice GGUFs under 1_7b/.
  The candidate bundle still stages with Q4_K_M (installable but the
  orchestrator's release-gate stays red until Q8_0 OmniVoice GGUFs are
  derived and pushed to elizaos/eliza-1-assets/27b/).

Verified:
- scripts/publish/test_orchestrator.py: 36/36 pass.
- scripts/test_backends_vast.py + scripts/test_vast_budget.py: 26/26 pass.
- bash scripts/train_vast.sh provision-and-train --registry-key qwen3.5-27b
  --dry-run: resolves gpu_target=b200-1x, world_size=1 (was b200-2x default).
- python -m scripts.publish.stage_base_v1_candidate --help: accepts --tier 27b.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: lalalune

packages/training/scripts/lib/vast.py

@@ -115,6 +115,13 @@
         "min_per_gpu_ram_gb": 130,
         "description": "1× H200 SXM (141 GB) — 9B SFT or 27B at low seq_len",
     },
+    "b200-1x": {
+        "gpu_names": ["B200"],
+        "num_gpus": 1,
+        # B200 = 180 GB HBM3e per GPU; gpu_ram>=170 robust to ECC reserve.
+        "min_per_gpu_ram_gb": 170,
+        "description": "1× NVIDIA B200 (≈183 GB) — qwen3.5-27b SFT default (130 GB budget @ seq=32k fits with headroom)",
+    },
     # ─── multi-GPU targets (27B+) ───
     "blackwell6000-2x": {
         # Both Server (S) and Workstation (WS) editions are 96 GB GDDR7

packages/training/scripts/publish/stage_base_v1_candidate.py

@@ -40,10 +40,26 @@
 REQUIRED_KERNELS_BY_TIER = {
     "0_6b": ["turboquant_q3", "qjl", "polarquant", "dflash"],
     "1_7b": ["turboquant_q4", "qjl", "polarquant", "dflash"],
+    # 27b matches eliza1_manifest.REQUIRED_KERNELS_BY_TIER["27b"] — adds
+    # turbo3_tcq on top of the base 1.7b set for long-context cache compression.
+    "27b": ["turboquant_q4", "qjl", "polarquant", "dflash", "turbo3_tcq"],
 }
 RAM_BUDGET_MB = {
     "0_6b": (2500, 3700),
     "1_7b": (4000, 5500),
+    # 27b matches publish/orchestrator.RAM_BUDGET_BY_TIER["27b"] — sized for
+    # 96GB+ Mac / high-VRAM desktop class hosts under the Q4_POLAR text bundle.
+    "27b": (24000, 32000),
+}
+# Per-tier upstream-Qwen3 substitute used by the lineage block and the
+# README/provenance prose. Falls back to "0.6B" for unknown tiers to match
+# the script's historical default behavior.
+QWEN3_PARAMS_BY_TIER = {
+    "0_6b": "0.6B",
+    "1_7b": "1.7B",
+    # The 27b cloud tier substitutes against Qwen3.5-27B (no Qwen3-27B
+    # variant exists upstream); the lineage block records the real base.
+    "27b": "27B",
 }
 TEXT_CTX = 32768
 
@@ -90,7 +106,7 @@ def download_asset(repo: str, remote_path: str, dest: Path) -> None:
 
 def main(argv: list[str] | None = None) -> int:
     ap = argparse.ArgumentParser(description=__doc__)
-    ap.add_argument("--tier", required=True, choices=("0_6b", "1_7b"))
+    ap.add_argument("--tier", required=True, choices=("0_6b", "1_7b", "27b"))
     ap.add_argument("--text-gguf", required=True, type=Path)
     ap.add_argument("--text-sidecar", type=Path, default=None,
                     help="The .eliza1.json sidecar for the text GGUF (quant block).")
@@ -171,9 +187,13 @@ def main(argv: list[str] | None = None) -> int:
             "sha256": drafter_sha,
             "source": args.drafter_source,
             "note": (
-                "Upstream Qwen3-0.6B GGUF used as the DFlash drafter for the "
-                "1.7B target; shares the Qwen3 BPE vocabulary so speculative "
-                "decoding is correct (modest acceptance — not a distilled drafter)."
+                # For 27b the canonical drafter is the Qwen3.5-aligned 0.6B
+                # distilled drafter (elizaos/eliza-1-drafter-0_6b-qwen3_5);
+                # for 0_6b/1_7b it's the upstream Qwen3 0.6B GGUF reused as-is.
+                f"DFlash drafter for the {QWEN3_PARAMS_BY_TIER.get(tier, '0.6B')} text target. "
+                "Shares the Qwen3.5/Qwen3 BPE vocabulary with the target so speculative "
+                "decoding is correct. See the drafter source repo for whether this "
+                "candidate is a distilled drafter or the upstream base GGUF (not distilled)."
             ),
         },
         "acceptanceWindow": None,
@@ -184,6 +204,15 @@ def main(argv: list[str] | None = None) -> int:
     # --- voice / asr / vad / cache from elizaos/eliza-1-assets/1_7b/ ---
     # The OmniVoice / Qwen3-ASR / Silero bytes are model-size-independent; the
     # assets repo only carries the 1_7b key, so reuse them under any tier.
+    #
+    # 27b caveat: eliza1_manifest.VOICE_QUANT_BY_TIER["27b"] == "Q8_0", so
+    # required_voice_artifacts_for_tier("27b") returns the Q8_0 names. This
+    # staging path still copies the Q4_K_M bytes (the only ones present in
+    # the assets repo today) — the orchestrator's voice-artifact gate will
+    # therefore fail in publish mode until Q8_0 OmniVoice GGUFs are derived
+    # and pushed to elizaos/eliza-1-assets/27b/. The candidate bundle is
+    # still installable on a runtime that can load Q4_K_M voice, but the
+    # release gate stays red. Track as a separate dependency.
     asset_map = [
         ("1_7b/tts/omnivoice-base-Q4_K_M.gguf", out / "tts" / "omnivoice-base-Q4_K_M.gguf"),
         ("1_7b/tts/omnivoice-tokenizer-Q4_K_M.gguf", out / "tts" / "omnivoice-tokenizer-Q4_K_M.gguf"),
@@ -288,9 +317,10 @@ def num(key: str) -> float | None:
         }, indent=2) + "\n")
 
     # --- lineage ---
+    params = QWEN3_PARAMS_BY_TIER.get(tier, "0.6B")
     lineage = {
         "text": M.LineageEntry(
-            base=f"{args.drafter_source.split('/')[0]}/Qwen3-{'1.7B' if tier=='1_7b' else '0.6B'} (SFT: APOLLO full-parameter; documented substitute for Qwen3.5-{'1.7B' if tier=='1_7b' else '0.6B'})",
+            base=f"{args.drafter_source.split('/')[0]}/Qwen3-{params} (SFT: APOLLO full-parameter; documented substitute for Qwen3.5-{params})",
             license="apache-2.0",
         ),
         "voice": M.LineageEntry(base="Serveurperso/OmniVoice-GGUF@361609388ae572a820d085185bbbe2a2aac4b30e", license="apache-2.0"),
@@ -319,15 +349,29 @@ def num(key: str) -> float | None:
             status="pass", at_commit="08032d57",
             report="packages/inference/verify/cuda-runtime-dispatch-evidence.json",
             device="NVIDIA GeForce RTX 5080 Laptop GPU (Blackwell, cc 12.0)",
-            caveat="cuda is not a tier-supported backend for 1_7b/0_6b — recorded as extra evidence",
+            # For 27b cuda is a tier-supported backend (per
+            # eliza1_manifest.SUPPORTED_BACKENDS_BY_TIER["27b"]); for 0_6b/1_7b
+            # it stays "extra evidence" — see the caveat tier-switch below.
+            caveat=(
+                "cuda is a tier-supported backend for 27b"
+                if tier == "27b"
+                else "cuda is not a tier-supported backend for 1_7b/0_6b — recorded as extra evidence"
+            ),
         ),
         "metal": M.KernelVerification(
             status="skipped", at_commit="08032d57", report="not-run",
             caveat="needs-hardware: no Apple/Metal device on the build host",
         ),
         "rocm": M.KernelVerification(
             status="skipped", at_commit="08032d57", report="not-applicable",
-            caveat="rocm is not a tier-supported backend for 1_7b/0_6b",
+            # rocm is a tier-supported backend for 27b but cannot be verified
+            # on this build host (no AMD GPU); 0_6b/1_7b don't list rocm as
+            # supported at all.
+            caveat=(
+                "rocm is a tier-supported backend for 27b but no AMD GPU on the build host (needs-hardware)"
+                if tier == "27b"
+                else "rocm is not a tier-supported backend for 1_7b/0_6b"
+            ),
         ),
     }
 
@@ -337,7 +381,7 @@ def num(key: str) -> float | None:
         "finetuned": True,
         "sourceModels": {
             "text": {
-                "repo": f"{args.drafter_source.split('/')[0]}/Qwen3-{'1.7B' if tier=='1_7b' else '0.6B'}",
+                "repo": f"{args.drafter_source.split('/')[0]}/Qwen3-{params}",
                 "convertedVia": "packages/inference/llama.cpp/convert_hf_to_gguf.py + scripts/optimize_for_eliza1.py (PolarQuant/QJL/TurboQuant)",
                 "note": "Fine-tuned (APOLLO full-parameter SFT) then optimized. Documented substitute for the not-yet-published Qwen3.5 base; NOT strictly base-v1 semantics — this is a finetuned candidate.",
             },
@@ -443,7 +487,7 @@ def _render_readme(
     optimized: bool,
     eval_results: dict[str, Any],
 ) -> str:
-    params = "1.7B" if tier == "1_7b" else "0.6B"
+    params = QWEN3_PARAMS_BY_TIER.get(tier, "0.6B")
     base_repo = f"{drafter_source.split('/')[0]}/Qwen3-{params}"
     if optimized:
         text_para = (

packages/training/scripts/train_vast.sh

@@ -248,6 +248,14 @@ case "$PIPELINE" in
         DEFAULT_GPU_TARGET="blackwell6000-1x"
         DEFAULT_FSDP_WORLD_SIZE=1
         ;;
+      qwen3.5-27b)
+        # Registry budget: 130 GB working set on a single 141 GB H200 or 183
+        # GB B200 (apollo_mini rank-1, grad ckpt, Liger CE, micro_batch=1
+        # seq=32k). B200-1x is the cheapest single-GPU fit (≈$3.8/hr × ~50h
+        # ≈ $190) and FSDP_WORLD_SIZE=1 matches the registry's extras block.
+        DEFAULT_GPU_TARGET="b200-1x"
+        DEFAULT_FSDP_WORLD_SIZE=1
+        ;;
       qwen3.6-27b)
         DEFAULT_GPU_TARGET="b200-2x"
         DEFAULT_FSDP_WORLD_SIZE=2