Promote first low-cost GGUF runtime evidence batch

docs/cli_reference.md

@@ -170,10 +170,10 @@ and global cache indices, so dense, MoE, tied-weight, quantized, and unknown
 architecture names do not need a runtime registry entry.
 
 Architecture-specific types remain only where the runtime selects different
-behavior. `lfm2` uses its legacy convolution-cache implementation. `gpt2` selects
-`Gpt_Model`, but Mobius's separate rank-4 key/value cache ABI does not match that
-runtime's rank-5 combined-cache contract, so config generation currently fails
-closed. `phi3`, `phimoe`, and `phi3small` retain their names only when their config selects
+behavior. `lfm2` uses its legacy convolution-cache implementation. `gpt2` uses the
+generic decoder because Mobius exports separate rank-4 key/value caches rather than
+the specialized `Gpt_Model` rank-5 combined-cache contract. `phi3`, `phimoe`, and
+`phi3small` retain their names only when their config selects
 LongRoPE, because the released generator uses those names to recompute caches when
 generation crosses the short-context threshold. Ordinary Phi-3-family graphs use
 `decoder`. Multimodal, audio, encoder-decoder, special-position-ID,
@@ -197,10 +197,11 @@ dense cache. The deferred state-manifest work is tracked by
 Each export also writes `runtime_compatibility.json`. Generic decoder metadata
 records the minimum runtime version and the latest stable release exercised by
 Mobius (0.15.2); it never emits the unreleased `decoder.state_groups` field.
-Generic config availability does not promote a GGUF runtime verdict: the only
-runtime-supported GGUF route remains the exact pinned SmolLM F16/CPU package, while
-SmolLM2 remains rejected because its GGUF padding-token metadata conflicts with the
-official pinned tokenizer.
+Generic config availability does not promote a GGUF runtime verdict. Runtime
+admission is restricted to exact artifacts named by the generated GGUF evidence
+matrix, which now includes GPT-2 and other independently verified decoder routes in
+addition to SmolLM. SmolLM2 remains rejected because its GGUF padding-token metadata
+conflicts with the official pinned tokenizer.
 
 #### Example
 

src/mobius/integrations/gguf/_arch_registry.py

@@ -659,7 +659,8 @@
     GGUFArchitectureSpec(
         gguf_arch="qwen3",
         model_type="qwen3",
-        tensor_map_recipe=("llama",),
+        tensor_map_recipe=("llama", "moe_qk_norm_extras"),
+        config_postprocessor="qwen3",
         runtime=Support.DEFERRED,
         reason=_RUNTIME_VALIDATION_PENDING,
     ),
@@ -1042,11 +1043,13 @@
         model_type="gpt2",
         tensor_map_recipe=("gpt2",),
         tensor_processor="gpt2",
-        runtime=Support.DEFERRED,
+        runtime=Support.SUPPORTED,
+        runtime_evidence_ids=("gpt2-q2-k-ort-genai-0.15.2",),
         quantized_import=Support.REJECTED,
         reason=(
-            _RUNTIME_VALIDATION_PENDING
-            + " Quantization preservation is rejected because canonical GPT-2 GGUF "
+            "Runtime support is restricted to the exact evidenced GPT-2 Q2_K artifact, "
+            "explicit-float CPU import route, pinned tokenizer, and ORT GenAI 0.15.2. "
+            "Quantization preservation is rejected because canonical GPT-2 GGUF "
             "projections must be transposed into graph order, and the current packed "
             "route cannot transpose values together with their scales and zero-points. "
             "Use keep_quantized=False for explicit float import."
@@ -1610,8 +1613,13 @@
         gguf_arch="starcoder2",
         model_type="starcoder2",
         tensor_map_recipe=("llama",),
-        runtime=Support.DEFERRED,
-        reason=_RUNTIME_VALIDATION_PENDING,
+        config_postprocessor="starcoder2",
+        runtime=Support.SUPPORTED,
+        runtime_evidence_ids=("tiny-starcoder2-q2-k-ort-genai-0.15.2",),
+        reason=(
+            "Runtime support is restricted to the exact evidenced tiny StarCoder2 Q2_K "
+            "artifact, explicit-float CPU import route, pinned tokenizer, and ORT GenAI 0.15.2."
+        ),
     ),
     GGUFArchitectureSpec(
         gguf_arch="stablelm",
@@ -1638,8 +1646,12 @@
         required_metadata=("attention.layer_norm_epsilon",),
         tensor_processor="llama",
         llama_qk_permute=True,
-        runtime=Support.DEFERRED,
-        reason=_RUNTIME_VALIDATION_PENDING,
+        runtime=Support.SUPPORTED,
+        runtime_evidence_ids=("tiny-olmo-q2-k-ort-genai-0.15.2",),
+        reason=(
+            "Runtime support is restricted to the exact evidenced tiny OLMo Q2_K artifact, "
+            "explicit-float CPU import route, pinned tokenizer, and ORT GenAI 0.15.2."
+        ),
     ),
     GGUFArchitectureSpec(
         gguf_arch="olmo2",
@@ -2174,11 +2186,13 @@
         tensor_map_recipe=("legacy_layernorm",),
         config_postprocessor="exact_legacy_gguf",
         required_metadata=("attention.layer_norm_epsilon", "attention.max_alibi_bias"),
-        runtime=Support.DEFERRED,
+        runtime=Support.SUPPORTED,
+        runtime_evidence_ids=("tiny-mpt-q2-k-ort-genai-0.15.2",),
         quantized_import=Support.REJECTED,
         reason=(
-            _RUNTIME_VALIDATION_PENDING
-            + " The admitted subset rejects learned positions, Q/K norms, KQV clipping, "
+            "Runtime support is restricted to the exact evidenced tiny MPT Q2_K artifact, "
+            "explicit-float portable graph, pinned tokenizer, and ORT GenAI 0.15.2. "
+            "The admitted subset rejects learned positions, Q/K norms, KQV clipping, "
             "AWQ activation scales, and inconsistent optional bias families. Quantization "
             "preservation is rejected because fused QKV must be split."
         ),
@@ -2190,11 +2204,13 @@
         tensor_map_recipe=("legacy_layernorm",),
         config_postprocessor="exact_legacy_gguf",
         required_metadata=("attention.layer_norm_epsilon", "use_parallel_residual"),
-        runtime=Support.DEFERRED,
+        runtime=Support.SUPPORTED,
+        runtime_evidence_ids=("pythia-70m-q2-k-ort-genai-0.15.2",),
         quantized_import=Support.REJECTED,
         reason=(
-            _RUNTIME_VALIDATION_PENDING
-            + " The admitted subset requires parallel residual MHA. Quantization "
+            "Runtime support is restricted to the exact evidenced Pythia-70M Q2_K artifact, "
+            "explicit-float portable graph, pinned tokenizer, and ORT GenAI 0.15.2. "
+            "The admitted subset requires parallel residual MHA. Quantization "
             "preservation is rejected because fused QKV rows must be split."
         ),
     ),
@@ -2336,11 +2352,13 @@
         tensor_map_recipe=("starcoder",),
         config_postprocessor="conventional_legacy",
         required_metadata=("attention.layer_norm_epsilon",),
-        runtime=Support.DEFERRED,
+        runtime=Support.SUPPORTED,
+        runtime_evidence_ids=("tiny-starcoder-q2-k-ort-genai-0.15.2",),
         quantized_import=Support.REJECTED,
         reason=(
-            _RUNTIME_VALIDATION_PENDING
-            + " Quantization preservation is rejected because StarCoder stores one fused "
+            "Runtime support is restricted to the exact evidenced tiny StarCoder Q2_K "
+            "artifact, explicit-float CPU import route, pinned tokenizer, and ORT GenAI "
+            "0.15.2. Quantization preservation is rejected because StarCoder stores one fused "
             "biased MQA projection that must be split for the graph."
         ),
     ),

src/mobius/integrations/gguf/_config_mapping.py

@@ -1533,6 +1533,27 @@ def _moe_postprocess(
     return dataclasses.replace(config, **updates)
 
 
+def _qwen3_postprocess(
+    config: ArchitectureConfig,
+    metadata: dict[str, Any],
+    model: Any = None,
+) -> ArchitectureConfig:
+    """Apply Qwen3's per-head Q/K normalization omitted by GGUF metadata."""
+    return dataclasses.replace(config, attn_qk_norm=True, attn_qk_norm_full=False)
+
+
+def _starcoder2_postprocess(
+    config: ArchitectureConfig,
+    metadata: dict[str, Any],
+    model: Any = None,
+) -> ArchitectureConfig:
+    """Restore StarCoder2's architecture-owned uniform sliding window."""
+    del metadata, model
+    if config.sliding_window is not None:
+        return config
+    return dataclasses.replace(config, sliding_window=4096)
+
+
 def _dbrx_postprocess(
     config: ArchitectureConfig,
     metadata: dict[str, Any],
@@ -4716,6 +4737,8 @@ def _qwen4exp_postprocess(
     "phimoe": _phimoe_postprocess,
     "pangu_embedded": _pangu_embedded_postprocess,
     "dense_sliding": _dense_sliding_postprocess,
+    "qwen3": _qwen3_postprocess,
+    "starcoder2": _starcoder2_postprocess,
     "gemma2": _gemma2_postprocess,
     "baichuan": _baichuan_postprocess,
     "chatglm": _chatglm_postprocess,
@@ -4764,15 +4787,17 @@ def _default_activation(model_type: str) -> str:
     # Most modern models use SiLU/Swish
     if model_type == "arcee":
         return "relu2"
+    if model_type in {"gpt2", "starcoder2"}:
+        # These architectures use tanh-approximate GELU by default. Their GGUF
+        # metadata omits that architecture-owned choice, so exact GELU is not equivalent.
+        return "gelu_pytorch_tanh"
     gelu_models = {
         "bert",
         "bloom",
-        "gpt2",
         "gpt_bigcode",
         "jais2",
         "kclgpt",
         "modernbert",
-        "starcoder2",
         "t5",
     }
     if model_type in gelu_models:

src/mobius/integrations/gguf/_config_mapping_test.py

@@ -1382,19 +1382,67 @@ def test_gemma_uses_gelu_pytorch_tanh(self, model_type: str) -> None:
 
         assert _default_activation(model_type) == "gelu_pytorch_tanh"
 
-    @pytest.mark.parametrize("model_type", ["gpt2", "bloom", "starcoder2", "t5"])
+    @pytest.mark.parametrize("model_type", ["bloom", "t5"])
     def test_gelu_models(self, model_type: str) -> None:
         from mobius.integrations.gguf._config_mapping import _default_activation
 
         assert _default_activation(model_type) == "gelu"
 
+    @pytest.mark.parametrize("model_type", ["gpt2", "starcoder2"])
+    def test_architecture_default_uses_tanh_approximate_gelu(self, model_type: str) -> None:
+        from mobius.integrations.gguf._config_mapping import _default_activation
+
+        assert _default_activation(model_type) == "gelu_pytorch_tanh"
+
     @pytest.mark.parametrize("model_type", ["llama", "qwen2", "mistral"])
     def test_silu_default(self, model_type: str) -> None:
         from mobius.integrations.gguf._config_mapping import _default_activation
 
         assert _default_activation(model_type) == "silu"
 
 
+def test_qwen3_qk_norm_weights_have_graph_consumers() -> None:
+    from mobius.integrations.gguf._config_mapping import gguf_to_config
+
+    tensor_names = [
+        "token_embd.weight",
+        "output.weight",
+        *[
+            f"blk.{layer}.{name}.weight"
+            for layer in range(2)
+            for name in ("attn_q_norm", "attn_k_norm")
+        ],
+    ]
+    config = gguf_to_config(_FakeDenseGGUF("qwen3", _dense_metadata("qwen3"), tensor_names))
+
+    assert config.attn_qk_norm is True
+    assert config.attn_qk_norm_full is False
+    module = registry.get(config.model_type)(config)
+    for layer in module.model.layers:
+        assert layer.self_attn.q_norm is not None
+        assert layer.self_attn.k_norm is not None
+
+
+@pytest.mark.parametrize(
+    ("metadata_window", "expected_window"),
+    [(None, 4096), (2048, 2048)],
+)
+def test_starcoder2_restores_architecture_sliding_window(
+    metadata_window: int | None,
+    expected_window: int,
+) -> None:
+    from mobius.integrations.gguf._config_mapping import gguf_to_config
+
+    metadata = _dense_metadata("starcoder2")
+    if metadata_window is not None:
+        metadata["starcoder2.attention.sliding_window"] = metadata_window
+    config = gguf_to_config(_FakeDenseGGUF("starcoder2", metadata, []))
+
+    assert config.sliding_window == expected_window
+    module = registry.get(config.model_type)(config)
+    assert module.model._sliding_window == expected_window
+
+
 class TestQwen35MtpBlockExclusion:
     """Qwen3.5/3.8 GGUF ``block_count`` includes trailing MTP (nextn) blocks.
 

src/mobius/integrations/gguf/_conventional_decoder_test.py

@@ -199,7 +199,10 @@ def test_conventional_decoder_support_and_runtime_verdicts_are_explicit() -> Non
     for architecture in _ARCHITECTURES:
         spec = get_arch_spec(architecture)
         assert spec.is_importable
-        assert spec.runtime is Support.DEFERRED
+        expected_runtime = (
+            Support.SUPPORTED if architecture == "starcoder" else Support.DEFERRED
+        )
+        assert spec.runtime is expected_runtime
         expected = (
             Support.SUPPORTED if architecture in supported_quantized else Support.REJECTED
         )

src/mobius/integrations/gguf/_docs.py

@@ -647,6 +647,12 @@ def render_document() -> str:
 
 ## Runtime evidence
 
+The first low-cost architecture batch promotes GPT-2, GPT-NeoX/Pythia, MPT, OLMo,
+StarCoder, and StarCoder2 using 334,238,976 bytes of GGUF payload and 346,825,051
+download bytes including tokenizer assets. Every route is explicit-float only. The
+network-free selection, budget, exclusions, and fail-closed candidate reasons are recorded in
+`testdata/evidence/gguf_low_cost_runtime_batch.json`.
+
 {_runtime_evidence_table()}
 
 Runtime support above is independent from tokenizer materialization support below.

src/mobius/integrations/gguf/_docs_test.py

@@ -39,7 +39,7 @@ def test_document_is_exact_generator_output() -> None:
 
 def test_document_is_concise_and_reason_coded() -> None:
     document = Path("docs/api/build_from_gguf.md").read_text(encoding="utf-8")
-    assert len(document.splitlines()) < 545
+    assert len(document.splitlines()) < 560
     assert "RUNTIME_EVIDENCE_PENDING" in document
     assert "qwen3.5-0.8b-q4-tokenizer" in document
     assert "qwen2.5-0.5b-instruct-q8-tokenizer" in document
@@ -143,6 +143,12 @@ def test_runtime_support_requires_structured_evidence() -> None:
         ("qwen2", ("qwen2.5-0.5b-instruct-q8-ort-genai-0.15.2",)),
         ("lfm2", ("lfm2-350m-f16-ort-genai-0.15.2",)),
         ("qwen35moe", ("qwen3.5-moe-0.87b-q2-k-ort-genai-0.15.2",)),
+        ("gpt2", ("gpt2-q2-k-ort-genai-0.15.2",)),
+        ("starcoder2", ("tiny-starcoder2-q2-k-ort-genai-0.15.2",)),
+        ("olmo", ("tiny-olmo-q2-k-ort-genai-0.15.2",)),
+        ("mpt", ("tiny-mpt-q2-k-ort-genai-0.15.2",)),
+        ("gptneox", ("pythia-70m-q2-k-ort-genai-0.15.2",)),
+        ("starcoder", ("tiny-starcoder-q2-k-ort-genai-0.15.2",)),
     ]
 
     pins = {pin.artifact_id for pin in MMPROJ_ARTIFACT_PINS}

src/mobius/integrations/gguf/_exact_legacy_decoder_test.py

@@ -154,7 +154,10 @@ def test_exact_legacy_config_tensor_and_graph_closure(architecture: str) -> None
             "openelm": "openelm",
         }[architecture]
     )
-    assert spec.runtime is Support.DEFERRED
+    expected_runtime = (
+        Support.SUPPORTED if architecture in {"gptneox", "mpt"} else Support.DEFERRED
+    )
+    assert spec.runtime is expected_runtime
     if architecture == "ernie4_5":
         assert config.rope_interleave
 

src/mobius/integrations/gguf/_quant_capabilities_test.py

@@ -167,10 +167,10 @@ def test_selected_real_artifacts_stay_within_global_budget() -> None:
     assert isinstance(policy, dict)
     assert isinstance(artifacts, list)
     assert isinstance(lossy, list)
-    assert len(artifacts) == 4
+    assert len(artifacts) == 10
     assert len(lossy) == 1
     selected = sum(int(record["size"]) for record in [*artifacts, *lossy])
-    assert selected == 2_390_132_320
+    assert selected == 2_724_371_296
     assert selected == policy["selected_artifact_bytes"]
     assert selected <= policy["max_selected_artifact_bytes"]
     assert lossy[0]["lfs_sha256"] == (

src/mobius/integrations/gguf/_route_census.py

@@ -86,6 +86,42 @@ class RecentPRDependency:
         "merged",
         "follow-up hardening for sharded GGUF evidence edge cases after PR #656",
     ),
+    RecentPRDependency(
+        672,
+        "Prove Qwen3.5 MoE GGUF explicit-float runtime route",
+        "merged",
+        "Qwen3.5 MoE runtime evidence is complete; exclude it from follow-up evidence batches",
+    ),
+    RecentPRDependency(
+        674,
+        "Keep Nemotron-H GGUF runtime fail-closed with pinned evidence",
+        "merged",
+        "Nemotron-H MoE blocker evidence is complete; exclude it from follow-up evidence batches",
+    ),
+    RecentPRDependency(
+        677,
+        "Promote first GGUF text architecture cohort",
+        "merged",
+        "the text architecture cohort is merged into the authoritative graph-route census",
+    ),
+    RecentPRDependency(
+        678,
+        "Promote generic GGUF projector cohort",
+        "merged",
+        "the generic projector cohort is merged into the authoritative projector census",
+    ),
+    RecentPRDependency(
+        679,
+        "Add exact Hunyuan V3 GGUF graph support",
+        "merged",
+        "Hunyuan V3 graph and MTP routes are merged into the authoritative route census",
+    ),
+    RecentPRDependency(
+        680,
+        "Promote first GGUF MoE architecture cohort",
+        "merged",
+        "the MoE architecture cohort is merged into the authoritative graph-route census",
+    ),
 )
 
 # Reviewed dispositions for routes that cannot advance until Mobius owns a new