invarlock 0.3.6__py3-none-any.whl → 0.3.7__py3-none-any.whl

invarlock/adapters/{hf_bert.py → hf_mlm.py}

@@ -1,15 +1,8 @@
 """
-HuggingFace BERT Model Adapter
+HuggingFace masked LM adapter.
 ==============================
 
-ModelAdapter implementation for HuggingFace BERT architecture models.
-
-This adapter provides BERT-specific integration including:
-- Support for BERT, RoBERTa, DistilBERT, and other BERT variants
-- Proper handling of bidirectional attention layers
-- Support for classification heads and pooling layers
-- Token type embeddings and position embeddings handling
-- Proper device-aware state serialization
+ModelAdapter implementation for HuggingFace masked language models.
 """
 
 from typing import Any
@@ -27,7 +20,7 @@ TensorType = torch.Tensor
 ModuleType = nn.Module
 
 
-class HF_BERT_Adapter(HFAdapterMixin, ModelAdapter):
+class HF_MLM_Adapter(HFAdapterMixin, ModelAdapter):
     """
     HuggingFace-specific ModelAdapter implementation for BERT models.
 
@@ -39,7 +32,7 @@ class HF_BERT_Adapter(HFAdapterMixin, ModelAdapter):
     - Device-aware state serialization
     """
 
-    name = "hf_bert"
+    name = "hf_mlm"
 
     def load_model(
         self, model_id: str, device: str = "auto", **kwargs: Any

invarlock/adapters/{hf_t5.py → hf_seq2seq.py}

@@ -1,11 +1,11 @@
 """
-HuggingFace T5 Model Adapter
-============================
+HuggingFace encoder-decoder adapter.
+===================================
 
-ModelAdapter implementation for HuggingFace T5 encoder-decoder models.
+ModelAdapter implementation for HuggingFace encoder-decoder (seq2seq) models.
 
-Loads AutoModelForSeq2SeqLM (e.g., t5-small/base/large) and exposes a minimal
-describe() sufficient for guard policies and reporting.
+Loads AutoModelForSeq2SeqLM and exposes a minimal describe() sufficient for
+guard policies and reporting.
 """
 
 from __future__ import annotations
@@ -25,10 +25,10 @@ TensorType = torch.Tensor
 ModuleType = nn.Module
 
 
-class HF_T5_Adapter(HFAdapterMixin, ModelAdapter):
-    """HuggingFace T5 adapter using AutoModelForSeq2SeqLM."""
+class HF_Seq2Seq_Adapter(HFAdapterMixin, ModelAdapter):
+    """HuggingFace encoder-decoder adapter using AutoModelForSeq2SeqLM."""
 
-    name = "hf_t5"
+    name = "hf_seq2seq"
 
     def load_model(  # type: ignore[override]
         self, model_id: str, device: str = "auto", **kwargs: Any
@@ -136,4 +136,4 @@ class HF_T5_Adapter(HFAdapterMixin, ModelAdapter):
         return super().restore(model, blob)
 
 
-__all__ = ["HF_T5_Adapter"]
+__all__ = ["HF_Seq2Seq_Adapter"]

invarlock/cli/adapter_auto.py

@@ -2,7 +2,8 @@
 Auto adapter resolution utilities.
 
 These helpers map a model identifier (HF directory or Hub ID) to a
-concrete built-in adapter name (hf_gpt2, hf_llama, hf_bert) without
+concrete built-in adapter name (hf_causal, hf_mlm, hf_seq2seq, hf_causal_onnx)
+without
 adding a hard dependency on Transformers.
 """
 
@@ -58,15 +59,15 @@ def _detect_quant_family_from_cfg(cfg: dict[str, Any]) -> str | None:
 
 
 def resolve_auto_adapter(
-    model_id: str | os.PathLike[str], default: str = "hf_gpt2"
+    model_id: str | os.PathLike[str], default: str = "hf_causal"
 ) -> str:
     """Resolve an appropriate built-in adapter name for a model.
 
     Heuristics:
       - Prefer local config.json (no network). Inspect `model_type` and
-        `architectures` to classify LLaMA/Mistral vs BERT vs GPT-like.
+        `architectures` to classify causal vs masked-LM vs seq2seq.
       - Fallback to simple name heuristics on the model_id string.
-      - Default to `hf_gpt2` when unsure.
+      - Default to `hf_causal` when unsure.
     """
     cfg = _read_local_hf_config(model_id)
     model_id_str = str(model_id)
@@ -77,32 +78,41 @@ def resolve_auto_adapter(
         if fam:
             return fam
         mt = str(c.get("model_type", "")).lower()
+        if bool(c.get("is_encoder_decoder", False)):
+            return "hf_seq2seq"
         archs = [str(a) for a in c.get("architectures", []) if isinstance(a, str)]
         arch_blob = " ".join(archs)
-        if (
-            mt in {"llama", "mistral", "qwen", "yi"}
-            or "Llama" in arch_blob
-            or "Mistral" in arch_blob
-        ):
-            return "hf_llama"
+        if "ConditionalGeneration" in arch_blob or "Seq2SeqLM" in arch_blob:
+            return "hf_seq2seq"
         # Treat masked-LM families as BERT-like
         if (
             mt in {"bert", "roberta", "distilbert", "albert", "deberta", "deberta-v2"}
             or "MaskedLM" in arch_blob
         ):
-            return "hf_bert"
-        # Generic causal LM
-        if "CausalLM" in arch_blob or mt in {
+            return "hf_mlm"
+        # Causal LM families (best-effort; structural validation happens in the adapter).
+        if "CausalLM" in arch_blob or "ForCausalLM" in arch_blob:
+            return "hf_causal"
+        if mt in {
+            "mistral",
+            "mixtral",
+            "qwen",
+            "qwen2",
+            "qwen2_moe",
+            "yi",
             "gpt2",
             "gpt_neox",
             "opt",
             "gptj",
-            "gptj8bit",
+            "phi",
+            "falcon",
+            "glm",
+            "deepseek",
         }:
-            return "hf_gpt2"
+            return "hf_causal"
         return None
 
-    # If local directory contains ONNX model files, prefer hf_onnx
+    # If local directory contains ONNX model files, prefer the ONNX causal adapter.
     try:
         p = Path(model_id)
         if p.exists() and p.is_dir():
@@ -114,7 +124,7 @@ def resolve_auto_adapter(
                 "encoder_model.onnx",
             ]
             if any((p / fname).exists() for fname in onnx_files):
-                return "hf_onnx"
+                return "hf_causal_onnx"
     except Exception:
         pass
 
@@ -134,10 +144,10 @@ def resolve_auto_adapter(
         k in lower_id for k in ["bnb", "bitsandbytes", "-4bit", "-8bit", "4bit", "8bit"]
     ):
         return "hf_bnb"
-    if any(k in lower_id for k in ["llama", "mistral", "qwen", "yi"]):
-        return "hf_llama"
+    if any(k in lower_id for k in ["t5", "bart"]):
+        return "hf_seq2seq"
     if any(k in lower_id for k in ["bert", "roberta", "albert", "deberta"]):
-        return "hf_bert"
+        return "hf_mlm"
     return default
 
 
@@ -148,7 +158,7 @@ def apply_auto_adapter_if_needed(cfg: Any) -> Any:
     """
     try:
         adapter = str(getattr(cfg.model, "adapter", ""))
-        if adapter.strip().lower() not in {"auto", "hf_auto", "auto_hf"}:
+        if adapter.strip().lower() not in {"auto", "auto_hf"}:
             return cfg
         model_id = str(getattr(cfg.model, "id", ""))
         resolved = resolve_auto_adapter(model_id)

invarlock/cli/app.py

@@ -105,7 +105,7 @@ Order: certify → report → run → plugins → doctor → version
 @app.command(
     name="certify",
     help=(
-        "Certify a subject model against a baseline and generate a safety certificate. "
+        "Certify a subject model against a baseline and generate an evaluation certificate. "
         "Use when you have two model snapshots and want pass/fail gating."
     ),
 )
@@ -116,6 +116,14 @@ def _certify_lazy(
     edited: str = typer.Option(
         ..., "--edited", "--subject", help="Subject model dir or Hub ID"
     ),
+    baseline_report: str | None = typer.Option(
+        None,
+        "--baseline-report",
+        help=(
+            "Reuse an existing baseline run report.json (skips baseline evaluation). "
+            "Must include stored evaluation windows (e.g., set INVARLOCK_STORE_EVAL_WINDOWS=1)."
+        ),
+    ),
     adapter: str = typer.Option(
         "auto", "--adapter", help="Adapter name or 'auto' to resolve"
     ),
@@ -139,12 +147,38 @@ def _certify_lazy(
     edit_config: str | None = typer.Option(
         None, "--edit-config", help="Edit preset to apply a demo edit (quant_rtn)"
     ),
+    edit_label: str | None = typer.Option(
+        None,
+        "--edit-label",
+        help=(
+            "Edit algorithm label for BYOE models. Use 'noop' for baseline, "
+            "'quant_rtn' etc. for built-in edits, 'custom' for pre-edited models."
+        ),
+    ),
+    quiet: bool = typer.Option(
+        False, "--quiet", "-q", help="Minimal output (suppress run/report detail)"
+    ),
+    verbose: bool = typer.Option(
+        False, "--verbose", "-v", help="Verbose output (include debug details)"
+    ),
+    banner: bool = typer.Option(
+        True, "--banner/--no-banner", help="Show header banner"
+    ),
+    style: str = typer.Option("audit", "--style", help="Output style (audit|friendly)"),
+    timing: bool = typer.Option(False, "--timing", help="Show timing summary"),
+    progress: bool = typer.Option(
+        True, "--progress/--no-progress", help="Show progress done messages"
+    ),
+    no_color: bool = typer.Option(
+        False, "--no-color", help="Disable ANSI colors (respects NO_COLOR=1)"
+    ),
 ):
     from .commands.certify import certify_command as _cert
 
     return _cert(
         source=source,
         edited=edited,
+        baseline_report=baseline_report,
         adapter=adapter,
         device=device,
         profile=profile,
@@ -153,6 +187,14 @@ def _certify_lazy(
         out=out,
         cert_out=cert_out,
         edit_config=edit_config,
+        edit_label=edit_label,
+        quiet=quiet,
+        verbose=verbose,
+        banner=banner,
+        style=style,
+        timing=timing,
+        progress=progress,
+        no_color=no_color,
     )
 
 
@@ -230,7 +272,7 @@ def _verify_typed(
     name="run",
     help=(
         "Execute an end-to-end run from a YAML config (edit + guards + reports). "
-        "Writes run artifacts and optionally a safety certificate."
+        "Writes run artifacts and optionally an evaluation certificate."
     ),
 )
 def _run_typed(
@@ -245,11 +287,24 @@ def _run_typed(
     ),
     out: str | None = typer.Option(None, "--out", help="Output directory override"),
     edit: str | None = typer.Option(None, "--edit", help="Edit kind (quant|mixed)"),
+    edit_label: str | None = typer.Option(
+        None,
+        "--edit-label",
+        help=(
+            "Edit algorithm label for BYOE models. Use 'noop' for baseline, "
+            "'quant_rtn' etc. for built-in edits, 'custom' for pre-edited models."
+        ),
+    ),
     tier: str | None = typer.Option(
         None,
         "--tier",
         help="Auto-tuning tier override (conservative|balanced|aggressive)",
     ),
+    metric_kind: str | None = typer.Option(
+        None,
+        "--metric-kind",
+        help="Primary metric kind override (ppl_causal|ppl_mlm|accuracy|etc.)",
+    ),
     probes: int | None = typer.Option(
         None, "--probes", help="Number of micro-probes (0=deterministic, >0=adaptive)"
     ),
@@ -270,6 +325,16 @@ def _run_typed(
     no_cleanup: bool = typer.Option(
         False, "--no-cleanup", help="Skip cleanup of temporary artifacts"
     ),
+    style: str | None = typer.Option(
+        None, "--style", help="Output style (audit|friendly)"
+    ),
+    progress: bool = typer.Option(
+        False, "--progress", help="Show progress done messages"
+    ),
+    timing: bool = typer.Option(False, "--timing", help="Show timing summary"),
+    no_color: bool = typer.Option(
+        False, "--no-color", help="Disable ANSI colors (respects NO_COLOR=1)"
+    ),
 ):
     from .commands.run import run_command as _run
 
@@ -279,13 +344,19 @@ def _run_typed(
         profile=profile,
         out=out,
         edit=edit,
+        edit_label=edit_label,
         tier=tier,
+        metric_kind=metric_kind,
         probes=probes,
         until_pass=until_pass,
         max_attempts=max_attempts,
         timeout=timeout,
         baseline=baseline,
         no_cleanup=no_cleanup,
+        style=style,
+        progress=progress,
+        timing=timing,
+        no_color=no_color,
     )