Feat/qwen3.6

atlas/cli/commands/model_registry.py

@@ -170,6 +170,13 @@ def _unsloth_qwen35_url(repo: str, file: str) -> str:
             f"{_UNSLOTH_QWEN35_COMMIT}/{file}")
 
 
+_UNSLOTH_QWEN36_COMMIT = "5cb35eb3dcbf52dbce5f87dbc64df6aaffadcace"
+
+def _unsloth_qwen36_url(file: str) -> str:
+    return (f"https://huggingface.co/unsloth/Qwen3.6-27B-MTP-GGUF/resolve/"
+            f"{_UNSLOTH_QWEN36_COMMIT}/{file}")
+
+
 # Single source of truth. Order: by tier (cpu → xlarge), then by quant.
 #
 # Truthful state today:
@@ -333,6 +340,34 @@ def _unsloth_qwen35_url(repo: str, file: str) -> str:
               "llama.cpp model only — G(x) verification will silently "
               "no-op (--no-lens to acknowledge). See PC-058 roadmap.",
     ),
+    Model(
+        name="Qwen3.6-27B-MTP-UD-Q4_K_XL",
+        tier="xlarge",
+        model_file="Qwen3.6-27B-UD-Q4_K_XL.gguf",
+        model_display="Qwen3.6 27B MTP (UD-Q4_K_XL)",
+        model_size_gb=16.7,
+        lens_status="no-artifacts",
+        download_url=_unsloth_qwen36_url("Qwen3.6-27B-UD-Q4_K_XL.gguf"),
+        sha256="4085665ee36d82a672a238a43f0e5643f2f0e39f2d7bd5d373f0ef10ecf53095",
+        license="Apache-2.0",
+        requires_hf_token=False,
+        asa_status="supported",
+        asa_artifact_files=["ast_edit_steering.gguf"],
+        asa_hf_repo="yogthos/atlas-asa-qwen3.6-27b-mtp-ud-q4_k_xl",
+        notes="Qwen3.6-27B with Multi-Token Prediction. Gated DeltaNet + "
+              "Gated Attention hybrid architecture, 5120-dim hidden. "
+              "MTP supported in llama.cpp since PR #22673. Requires "
+              "--spec-type draft-mtp --spec-draft-n-max 2 flags. "
+              "ASA control vector (ast_edit_steering.gguf, layer 27, "
+              "5120-dim) published at "
+              "https://huggingface.co/yogthos/atlas-asa-qwen3.6-27b-mtp-ud-q4_k_xl. "
+              "Lens C(x) cost_field.pt trained locally (5120-dim, "
+              "200-epoch contrastive) but not yet published on HF, and "
+              "the G(x) classifier half (gx_*.json) is untrained — G(x) "
+              "will silently no-op until both lens halves ship via "
+              "`atlas lens publish`. macOS Metal: MTP disabled by default "
+              "(ATLAS_ENABLE_MTP=0); see llama.cpp #23011/#23752.",
+    ),
     Model(
         name="gemma-4-12b-it-Q4_K_M",
         tier="medium",

docs/models/qwen3.6.md

@@ -0,0 +1,127 @@
+# Qwen3.6-27B-MTP on ATLAS (macOS Metal)
+
+## Model specs
+
+- Architecture: Gated DeltaNet + Gated Attention hybrid, 5120-dim hidden, 64 layers
+- MTP: Multi-Token Prediction (inline draft head, `--spec-type draft-mtp`)
+- Context: 262K native, 32K recommended for 64GB M1 Max
+- Quantization: UD-Q4_K_XL (19 GB on disk, ~24 GB VRAM at 32K context)
+- License: Apache 2.0
+
+## What works
+
+- Base inference via Metal-accelerated llama.cpp
+- Lens C(x) cost-field scoring (5120-dim, 200-epoch contrastive training)
+- Full V3 pipeline (budget forcing, plan search, candidate selection, refinement loop)
+
+## What doesn't
+
+- **MTP speculative decoding**: crashes during draft context init (`GGML_ASSERT missing result_norm/result_embd`) and has net throughput loss on Metal even when it doesn't crash (llama.cpp #23752). Disabled by default (`ATLAS_ENABLE_MTP=0`).
+- **ASA steering vectors**: no vector trained for Qwen3.6 residuals. Would need `atlas asa build` against this model.
+- **e2e_smoke in atlas doctor**: 300 tokens × ~5 tok/s on M1 Max exceeds the 60s hard timeout. The model itself works fine for interactive use.
+
+## Fresh clone setup
+
+```bash
+# 1. Clone and configure
+git clone https://github.com/itigges22/ATLAS
+cd ATLAS
+cp .env.example .env
+
+# Edit .env:
+#   ATLAS_MODEL_FILE=Qwen3.6-27B-UD-Q4_K_XL.gguf
+#   ATLAS_MODEL_NAME=Qwen3.6-27B-MTP-UD-Q4_K_XL
+#   ATLAS_ENABLE_MTP=0
+
+# 2. Download the model (17.9 GB)
+hf download unsloth/Qwen3.6-27B-MTP-GGUF \
+  --include "*UD-Q4_K_XL*" \
+  --local-dir ./models
+
+# 3. Build native Metal llama-server
+./scripts/atlas-setup-macos.sh
+
+# 4. Start the native server (keep this terminal open)
+source .env
+./scripts/atlas-llama-macos.sh
+
+# 5. Start Docker services (new terminal)
+docker compose -f docker-compose.yml -f docker-compose.macos.yml up -d
+
+# 6. Verify
+export ATLAS_MODEL_FILE=Qwen3.6-27B-UD-Q4_K_XL.gguf
+atlas doctor
+```
+
+## Lens training (optional, ~2h on CPU)
+
+```bash
+# Download pre-computed 5120-dim embeddings
+curl -sL "https://huggingface.co/datasets/itigges22/ATLAS/resolve/main/embeddings/training_embeddings_5120d.json" \
+  -o ./geometric-lens/geometric_lens/models/qwen36-27b/training_embeddings_5120d.json
+
+# Copy data into the running geometric-lens container
+docker cp ./geometric-lens/geometric_lens/models/qwen36-27b/training_embeddings_5120d.json \
+  atlas-geometric-lens-1:/tmp/
+
+# Train C(x) inside the container (torch is pre-installed there)
+docker exec -w /app atlas-geometric-lens-1 python3 -c "
+import json, os
+from geometric_lens.training import train_cost_field, save_cost_field
+
+with open('/tmp/training_embeddings_5120d.json') as f:
+    data = json.load(f)
+data['labels'] = [1 if l == 'PASS' else 0 for l in data['labels']]
+print(f'Loaded {len(data[\"embeddings\"])} embeddings, dim={data[\"dim\"]}')
+
+result = train_cost_field(data, epochs=200)
+test_auc = result.get('best_test_auc') or result.get('final_test_auc') or 0.0
+print(f'Test AUC: {test_auc:.4f}')
+
+os.makedirs('/tmp/qwen36-27b', exist_ok=True)
+cost_path = save_cost_field(result['model'], save_dir='/tmp/qwen36-27b')
+print(f'Saved: {cost_path}')
+"
+
+# Copy trained artifact back to host
+docker cp atlas-geometric-lens-1:/tmp/qwen36-27b/cost_field.pt \
+  ./geometric-lens/geometric_lens/models/qwen36-27b/
+
+# Doctor requires metric_tensor.pt on disk (runtime uses XGBoost, not this file)
+touch ./geometric-lens/geometric_lens/models/qwen36-27b/metric_tensor.pt
+
+# Verify Lens
+export ATLAS_LENS_MODELS=./geometric-lens/geometric_lens/models/qwen36-27b
+atlas doctor
+```
+
+## Expected doctor output
+
+```
+  ✓ health/llama                      ok
+  ✓ model_file                        Qwen3.6-27B-UD-Q4_K_XL.gguf (16.7 GB)
+  ✓ lens_weights                      cost_field.pt + metric_tensor.pt (after training)
+  ⚠ asa_steering                      expected — no vector trained for Qwen3.6
+  ⚠ tier_constraints                  disk free — Qwen3.6 is 17.9 GB
+  ✗ e2e_smoke                         timeout — 27B on M1 Max is slow, not a bug
+
+  19 passed, 2 warnings, 1 failed, 1 skipped
+  ATLAS install has failures — re-run with -v for detail.
+```
+
+## Registry entry
+
+The model is registered in `atlas/cli/commands/model_registry.py` as:
+
+- `name`: `Qwen3.6-27B-MTP-UD-Q4_K_XL`
+- `tier`: `xlarge`
+- `lens_status`: `no-artifacts` (until artifacts are published to HF)
+- `download_url`: `None` (manual download via `hf download`)
+
+## Key files
+
+- `atlas/cli/commands/model_registry.py` — registry entry
+- `inference/Dockerfile.v31` — llama.cpp SHA `25558268` (MTP merge)
+- `inference/patches/expose-hidden-states.patch` — regenerated for new upstream
+- `scripts/atlas-llama-macos.sh` — MTP flag gating via `ATLAS_ENABLE_MTP`
+- `geometric-lens/geometric_lens/models/qwen36-27b/` — Lens artifacts

geometric-lens/geometric_lens/models/qwen36-27b/train_cx.py

@@ -0,0 +1,30 @@
+#!/usr/bin/env python3
+"""Train C(x) cost field from pre-computed 5120-dim embeddings."""
+import json, os, sys, time
+
+path = "/app/geometric_lens/models/qwen36-27b/training_embeddings_5120d.json"
+with open(path) as f:
+    data = json.load(f)
+
+sys.stderr.write(f"Loaded {len(data['embeddings'])} embeddings, dim={data['dim']}\n")
+labels = [1 if l == "PASS" else 0 for l in data["labels"]]
+data["labels"] = labels
+sys.stderr.write(f"Labels: {sum(labels)} PASS, {len(labels) - sum(labels)} FAIL\n")
+sys.stderr.flush()
+
+from geometric_lens.training import train_cost_field, save_cost_field
+
+start = time.time()
+result = train_cost_field(data, epochs=200, lr=1e-3, margin=1.0)
+elapsed = time.time() - start
+
+test_auc = result.get("best_test_auc") or result.get("final_test_auc") or 0.0
+train_auc = result.get("final_train_auc") or 0.0
+sys.stderr.write(f"Train AUC: {train_auc:.4f}  |  Test AUC: {test_auc:.4f}  |  Time: {elapsed:.0f}s\n")
+sys.stderr.flush()
+
+artifact_dir = "/app/geometric_lens/models/qwen36-27b"
+cost_path = save_cost_field(result["model"], save_dir=artifact_dir)
+sys.stderr.write(f"Saved: {cost_path}\n")
+sys.stderr.flush()
+print("DONE")

inference/Dockerfile.v31

@@ -62,7 +62,7 @@ COPY patches/ /patches/
 #   1. There's a llama.cpp release with a feature/fix we want
 #   2. The PC-202 patch has been regenerated against the new upstream
 # Override at build time: --build-arg LLAMA_CPP_REV=<sha>
-ARG LLAMA_CPP_REV=2e97c5f96f9fe2bb26f794a348e05d7a1c74baa1
+ARG LLAMA_CPP_REV=255582687b8dd211fdbc582e43ab842491554e94
 # #117: Split fetch -> patch -> build into three RUN layers so a
 # LLAMA_CPP_REV bump or a patches/ change invalidates only the affected
 # layer + everything below, not the whole giant step. The build is the