model-unfolder 0.2.3__tar.gz → 0.2.4__tar.gz

{model_unfolder-0.2.3 → model_unfolder-0.2.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: model-unfolder
-Version: 0.2.3
+Version: 0.2.4
 Summary: Unfold any HuggingFace transformer into an interactive architecture diagram, inline in Jupyter.
 Author: model-unfolder contributors
 License: Apache-2.0
@@ -110,7 +110,8 @@ Param estimates are close to published numbers — DeepSeek-V3 reports `~675B (~
 |---|---|
 | DeepSeek | DeepSeek-V2, DeepSeek-V3, Kimi K2 |
 | Llama | Llama 3 / 3.1 / 3.2 / 3.3 |
-| Mistral | v0.1 / 0.2 / 0.3 | 
+| Mistral | Mistral 7B, Mixtral 8x7B / 8x22B, Mistral Medium 3.5 |
+| Qwen | Qwen2 / 2.5, Qwen2-MoE, Qwen3, Qwen3-MoE, Qwen3.5 / 3.6 (hybrid linear+full attn) |
 | Gemma | Gemma 3, Gemma 4 (31B, E2B, E4B) |
 
 ### Diffusors

{model_unfolder-0.2.3 → model_unfolder-0.2.4}/README.md

@@ -86,7 +86,8 @@ Param estimates are close to published numbers — DeepSeek-V3 reports `~675B (~
 |---|---|
 | DeepSeek | DeepSeek-V2, DeepSeek-V3, Kimi K2 |
 | Llama | Llama 3 / 3.1 / 3.2 / 3.3 |
-| Mistral | v0.1 / 0.2 / 0.3 | 
+| Mistral | Mistral 7B, Mixtral 8x7B / 8x22B, Mistral Medium 3.5 |
+| Qwen | Qwen2 / 2.5, Qwen2-MoE, Qwen3, Qwen3-MoE, Qwen3.5 / 3.6 (hybrid linear+full attn) |
 | Gemma | Gemma 3, Gemma 4 (31B, E2B, E4B) |
 
 ### Diffusors

model_unfolder-0.2.4/model_unfolder/adapters/transformer/families/__init__.py

@@ -0,0 +1,13 @@
+"""Transformer model-family adapters.
+
+These modules translate family-specific HuggingFace config dialects into the
+shared transformer IR pieces in ``model_unfolder.adapters.transformer``.
+"""
+
+from . import deepseek, fallback, gemma4, llama, mistral, qwen
+
+# Order matters: more specific adapters first.
+# ``gemma4`` must run before ``llama`` (which also matches ``gemma`` model_type).
+# ``mistral`` and ``qwen`` must run before ``llama``.
+# ``fallback`` always matches — must be last.
+ADAPTERS = [deepseek, gemma4, mistral, qwen, llama, fallback]

model_unfolder-0.2.4/model_unfolder/adapters/transformer/families/fallback.py

@@ -0,0 +1,192 @@
+"""Generic fallback adapter for unknown transformer architectures.
+
+Registered last in ADAPTERS so it only fires when no specific family matches.
+It tries a broad set of field-name aliases used across different codebases, so
+most decoder-style configs parse correctly even without a dedicated adapter.
+
+Emits a warning in ModelIR.warnings for every gap it detects — unknown
+model_type, unrecognised layer_type strings, missing critical fields, etc.
+"""
+from __future__ import annotations
+
+from typing import Any
+
+from ....ir import AttentionSpec, FFNSpec, ModelIR
+from ..assembly import decoder_extras, decoder_layer
+from ..common import architecture_name, get_config_value as _g, model_name
+
+
+# ---------------------------------------------------------------------------
+# Multi-alias field resolver
+# ---------------------------------------------------------------------------
+
+_ALIASES: dict[str, list[str]] = {
+    "num_hidden_layers":       ["num_hidden_layers", "n_layers", "num_layers", "n_layer",
+                                 "num_blocks", "n_blocks"],
+    "num_attention_heads":     ["num_attention_heads", "n_heads", "num_heads", "n_head",
+                                 "num_q_heads"],
+    "num_key_value_heads":     ["num_key_value_heads", "n_kv_heads", "num_kv_heads",
+                                 "num_key_heads"],
+    "hidden_size":             ["hidden_size", "d_model", "n_embd", "model_dim",
+                                 "embed_dim", "dim"],
+    "intermediate_size":       ["intermediate_size", "ffn_dim", "mlp_dim", "inner_dim",
+                                 "ffn_hidden_size", "feed_forward_proj_dim"],
+    "hidden_act":              ["hidden_act", "activation_function", "hidden_activation",
+                                 "act_fn", "activation"],
+    "vocab_size":              ["vocab_size", "n_vocab", "padded_vocab_size"],
+    "max_position_embeddings": ["max_position_embeddings", "max_seq_len", "n_positions",
+                                 "context_length", "max_seq_length", "seq_length"],
+    "sliding_window":          ["sliding_window", "attention_window", "window_size"],
+    "num_experts":             ["num_experts", "num_local_experts", "n_experts"],
+    "num_experts_per_tok":     ["num_experts_per_tok", "top_k_experts", "top_k",
+                                 "num_selected_experts"],
+    "num_shared_experts":      ["num_shared_experts", "n_shared_experts"],
+    "moe_intermediate_size":   ["moe_intermediate_size", "expert_intermediate_size",
+                                 "expert_hidden_size", "ffn_dim_multiplier"],
+    "head_dim":                ["head_dim", "d_head", "head_size", "kv_channels"],
+    "tie_word_embeddings":     ["tie_word_embeddings", "tie_embeddings",
+                                 "tie_word_embedding_weights"],
+}
+
+
+def _resolve(cfg: Any, canonical: str, default=None):
+    """Try every known alias for a field, return the first hit."""
+    for alias in _ALIASES.get(canonical, [canonical]):
+        val = _g(cfg, alias)
+        if val is not None:
+            return val
+    return default
+
+
+def _unwrap_text(cfg: Any) -> Any:
+    """If a multimodal wrapper hides the LM config under a sub-key, unwrap it."""
+    for key in ("text_config", "language_config", "llm_config", "text_model_config"):
+        sub = _g(cfg, key)
+        if isinstance(sub, dict) and sub.get("num_hidden_layers") or (
+            hasattr(sub, "num_hidden_layers") and sub is not None
+        ):
+            return sub
+    return cfg
+
+
+# ---------------------------------------------------------------------------
+# Adapter interface
+# ---------------------------------------------------------------------------
+
+def matches(_cfg: Any) -> bool:
+    return True  # always fires — must be registered last
+
+
+def parse(cfg: Any) -> ModelIR:
+    warnings: list[str] = []
+
+    model_type = (_g(cfg, "model_type") or "unknown").lower()
+    arch_name  = architecture_name(cfg, "unknown")
+    warnings.append(
+        f"No dedicated adapter for model_type={model_type!r} / arch={arch_name!r}. "
+        "Parsed with generic fallback — some details may be approximate."
+    )
+
+    text_cfg = _unwrap_text(cfg)
+    if text_cfg is not cfg:
+        warnings.append(
+            "Config fields read from nested text_config sub-key (multimodal wrapper detected)."
+        )
+
+    num_layers  = _resolve(text_cfg, "num_hidden_layers", 0)
+    num_heads   = _resolve(text_cfg, "num_attention_heads", 0)
+    num_kv_heads = _resolve(text_cfg, "num_key_value_heads") or num_heads
+    hidden_size  = _resolve(text_cfg, "hidden_size", 0)
+    head_dim     = _resolve(text_cfg, "head_dim") or (hidden_size // num_heads if num_heads else None)
+    activation   = (_resolve(text_cfg, "hidden_act") or "silu").lower()
+    sliding_window = _resolve(text_cfg, "sliding_window")
+    layer_types  = _g(text_cfg, "layer_types") or []
+
+    if not num_layers:
+        warnings.append("Could not determine num_hidden_layers — layer list will be empty.")
+    if not hidden_size:
+        warnings.append("Could not determine hidden_size.")
+
+    if num_kv_heads == num_heads:
+        attn_kind = "mha"
+    elif num_kv_heads == 1:
+        attn_kind = "mqa"
+    else:
+        attn_kind = "gqa"
+
+    num_experts         = _resolve(text_cfg, "num_experts", 0)
+    num_experts_per_tok = _resolve(text_cfg, "num_experts_per_tok", 0)
+    num_shared_experts  = _resolve(text_cfg, "num_shared_experts", 0)
+    moe_intermediate_size = _resolve(text_cfg, "moe_intermediate_size", 0)
+    intermediate_size   = _resolve(text_cfg, "intermediate_size", 0) or moe_intermediate_size
+    is_moe = bool(num_experts)
+
+    unknown_layer_types: set[str] = set()
+
+    layers = []
+    for i in range(num_layers):
+        layer_type = layer_types[i] if i < len(layer_types) else "full_attention"
+
+        if layer_type in ("full_attention", "causal", ""):
+            mask, win = "causal", None
+        elif "sliding" in layer_type:
+            mask, win = "sliding", sliding_window
+        elif layer_type == "linear_attention":
+            mask, win = "causal", None
+            attn_kind = "linear"
+        else:
+            unknown_layer_types.add(layer_type)
+            mask, win = "causal", None
+
+        attn = AttentionSpec(
+            kind=attn_kind,
+            num_heads=num_heads,
+            num_kv_heads=num_kv_heads,
+            head_dim=head_dim,
+            mask=mask,
+            window_size=win,
+        )
+
+        if is_moe:
+            ffn = FFNSpec(
+                kind="moe",
+                activation=activation,
+                intermediate_size=intermediate_size,
+                gated=True,
+                num_experts=num_experts,
+                num_experts_per_tok=num_experts_per_tok,
+                num_shared_experts=num_shared_experts,
+                expert_intermediate_size=moe_intermediate_size or intermediate_size,
+            )
+        else:
+            ffn = FFNSpec(
+                kind="dense",
+                activation=activation,
+                intermediate_size=intermediate_size,
+                gated=True,
+            )
+
+        layers.append(decoder_layer(i, attn, ffn, hidden_size))
+
+    for lt in sorted(unknown_layer_types):
+        warnings.append(
+            f"Unrecognised layer_type={lt!r} — treated as standard causal attention. "
+            "Add a dedicated adapter to handle this correctly."
+        )
+
+    vocab_size = _resolve(text_cfg, "vocab_size", 0) or _resolve(cfg, "vocab_size", 0)
+    tie_word_embeddings = bool(
+        _resolve(text_cfg, "tie_word_embeddings", _resolve(cfg, "tie_word_embeddings", False))
+    )
+
+    return ModelIR(
+        name=model_name(cfg, arch_name),
+        architecture=arch_name,
+        vocab_size=vocab_size,
+        hidden_size=hidden_size,
+        max_position_embeddings=_resolve(text_cfg, "max_position_embeddings"),
+        tie_word_embeddings=tie_word_embeddings,
+        layers=layers,
+        extras=decoder_extras(vocab_size, hidden_size, tie_word_embeddings),
+        warnings=warnings,
+    )

{model_unfolder-0.2.3 → model_unfolder-0.2.4}/model_unfolder/adapters/transformer/families/llama.py

@@ -1,4 +1,4 @@
-"""Adapter for Llama, Mistral, Qwen, and similar GQA/MHA dense models."""
+"""Adapter for Llama, Phi, and similar GQA/MHA dense models."""
 from __future__ import annotations
 
 from typing import Any
@@ -8,26 +8,30 @@ from ..assembly import decoder_extras, decoder_layer
 from ..common import architecture_name, get_config_value as _g, model_name
 
 
-_FAMILIES = {"llama", "mistral", "qwen2", "qwen3", "phi3", "gemma"}
+_FAMILIES = {"llama", "phi3", "gemma"}
 
 
 def matches(cfg: Any) -> bool:
-    arches = _g(cfg, "architectures") or []
-    model_type = _g(cfg, "model_type", "")
-    for arch in arches:
-        if any(fam in arch.lower() for fam in ("llama", "mistral", "qwen", "phi3")):
-            return True
+    model_type = (_g(cfg, "model_type") or "").lower()
     if model_type in _FAMILIES:
         return True
-    return False
+    arches = _g(cfg, "architectures") or []
+    return any(
+        any(fam in a.lower() for fam in ("llama", "phi3"))
+        for a in arches
+    )
 
 
 def parse(cfg: Any) -> ModelIR:
+    arch_name = architecture_name(cfg, "llama")
+
     num_layers = _g(cfg, "num_hidden_layers", 0)
     num_heads = _g(cfg, "num_attention_heads", 0)
     num_kv_heads = _g(cfg, "num_key_value_heads", num_heads)
     hidden_size = _g(cfg, "hidden_size", 0)
     head_dim = _g(cfg, "head_dim") or (hidden_size // num_heads if num_heads else None)
+    intermediate_size = _g(cfg, "intermediate_size", 0)
+    activation = (_g(cfg, "hidden_act", "silu") or "silu").lower()
 
     if num_kv_heads == num_heads:
         attn_kind = "mha"
@@ -40,11 +44,6 @@ def parse(cfg: Any) -> ModelIR:
     sliding_pattern = _g(cfg, "sliding_window_pattern")
     layer_types = _g(cfg, "layer_types")
 
-    intermediate_size = _g(cfg, "intermediate_size", 0)
-    activation = (_g(cfg, "hidden_act", "silu") or "silu").lower()
-
-    arch_name = architecture_name(cfg, "llama")
-
     layers = []
     for i in range(num_layers):
         if layer_types and i < len(layer_types):

model_unfolder-0.2.4/model_unfolder/adapters/transformer/families/mistral.py

@@ -0,0 +1,119 @@
+"""Adapter for the Mistral model family.
+
+Covers:
+* Mistral 7B / Mistral Small / Ministral  — flat GQA config (model_type: "mistral")
+* Mixtral 8x7B / 8x22B                   — sparse MoE (model_type: "mixtral")
+* Mistral Medium 3.5 / Pixtral-class      — multimodal wrapper with text nested
+                                            under text_config (model_type: "mistral3")
+"""
+from __future__ import annotations
+
+from typing import Any
+
+from ....ir import AttentionSpec, FFNSpec, ModelIR
+from ..assembly import decoder_extras, decoder_layer
+from ..common import architecture_name, get_config_value as _g, model_name
+
+
+_FLAT_TYPES  = {"mistral", "mixtral"}
+_WRAPPED_TYPES = {"mistral3", "ministral3"}
+_ALL_TYPES = _FLAT_TYPES | _WRAPPED_TYPES
+
+
+def matches(cfg: Any) -> bool:
+    model_type = (_g(cfg, "model_type") or "").lower()
+    if model_type in _ALL_TYPES:
+        return True
+    arches = _g(cfg, "architectures") or []
+    return any("mistral" in a.lower() or "mixtral" in a.lower() for a in arches)
+
+
+def parse(cfg: Any) -> ModelIR:
+    text_cfg = _text_config(cfg)
+    arch_name = architecture_name(cfg, "mistral")
+
+    num_layers    = _g(text_cfg, "num_hidden_layers", 0)
+    num_heads     = _g(text_cfg, "num_attention_heads", 0)
+    num_kv_heads  = _g(text_cfg, "num_key_value_heads", num_heads)
+    hidden_size   = _g(text_cfg, "hidden_size", 0)
+    head_dim      = _g(text_cfg, "head_dim") or (hidden_size // num_heads if num_heads else None)
+    activation    = (_g(text_cfg, "hidden_act", "silu") or "silu").lower()
+
+    if num_kv_heads == num_heads:
+        attn_kind = "mha"
+    elif num_kv_heads == 1:
+        attn_kind = "mqa"
+    else:
+        attn_kind = "gqa"
+
+    sliding_window  = _g(text_cfg, "sliding_window")
+    sliding_pattern = _g(text_cfg, "sliding_window_pattern")
+
+    # MoE fields (Mixtral uses num_local_experts; generic fallback to num_experts)
+    num_experts        = _g(text_cfg, "num_local_experts") or _g(text_cfg, "num_experts") or 0
+    num_experts_per_tok = _g(text_cfg, "num_experts_per_tok") or 0
+    intermediate_size  = _g(text_cfg, "intermediate_size", 0)
+    is_moe = bool(num_experts)
+
+    layers = []
+    for i in range(num_layers):
+        if sliding_pattern and sliding_window:
+            mask = "sliding" if (i % sliding_pattern) != (sliding_pattern - 1) else "causal"
+            win  = sliding_window if mask == "sliding" else None
+        elif sliding_window:
+            mask, win = "sliding", sliding_window
+        else:
+            mask, win = "causal", None
+
+        attn = AttentionSpec(
+            kind=attn_kind,
+            num_heads=num_heads,
+            num_kv_heads=num_kv_heads,
+            head_dim=head_dim,
+            mask=mask,
+            window_size=win,
+        )
+
+        if is_moe:
+            ffn = FFNSpec(
+                kind="moe",
+                activation=activation,
+                intermediate_size=intermediate_size,
+                gated=True,
+                num_experts=num_experts,
+                num_experts_per_tok=num_experts_per_tok,
+                expert_intermediate_size=intermediate_size,
+            )
+        else:
+            ffn = FFNSpec(
+                kind="dense",
+                activation=activation,
+                intermediate_size=intermediate_size,
+                gated=True,
+            )
+
+        layers.append(decoder_layer(i, attn, ffn, hidden_size))
+
+    vocab_size = _g(text_cfg, "vocab_size", 0) or _g(cfg, "vocab_size", 0)
+    tie_word_embeddings = bool(
+        _g(text_cfg, "tie_word_embeddings", _g(cfg, "tie_word_embeddings", False))
+    )
+    return ModelIR(
+        name=model_name(cfg, arch_name),
+        architecture=arch_name,
+        vocab_size=vocab_size,
+        hidden_size=hidden_size,
+        max_position_embeddings=_g(text_cfg, "max_position_embeddings"),
+        tie_word_embeddings=tie_word_embeddings,
+        layers=layers,
+        extras=decoder_extras(vocab_size, hidden_size, tie_word_embeddings),
+    )
+
+
+def _text_config(cfg: Any) -> Any:
+    model_type = (_g(cfg, "model_type") or "").lower()
+    if model_type in _WRAPPED_TYPES:
+        sub = _g(cfg, "text_config")
+        if sub is not None:
+            return sub
+    return cfg

model_unfolder-0.2.4/model_unfolder/adapters/transformer/families/qwen.py

@@ -0,0 +1,144 @@
+"""Adapter for the Qwen model family.
+
+Covers:
+* Qwen2 / Qwen2.5 (dense)     — model_type: "qwen2"
+* Qwen2-MoE                   — model_type: "qwen2_moe"
+* Qwen3 (dense)               — model_type: "qwen3"
+* Qwen3-MoE / Qwen2.5-Max     — model_type: "qwen3_moe"
+* Qwen3.5 / Qwen3.6 (hybrid)  — model_type: "qwen3_5_moe", text nested under text_config.
+                                 Alternates linear (SSM-style) and full-context attention layers.
+"""
+from __future__ import annotations
+
+from typing import Any
+
+from ....ir import AttentionSpec, FFNSpec, ModelIR
+from ..assembly import decoder_extras, decoder_layer
+from ..common import architecture_name, get_config_value as _g, model_name
+
+
+_FLAT_TYPES    = {"qwen2", "qwen2_moe", "qwen3", "qwen3_moe"}
+_WRAPPED_TYPES = {"qwen3_5_moe", "qwen3_5_moe_text"}
+_MODEL_TYPES   = _FLAT_TYPES | _WRAPPED_TYPES
+
+
+def matches(cfg: Any) -> bool:
+    model_type = (_g(cfg, "model_type") or "").lower()
+    if model_type in _MODEL_TYPES:
+        return True
+    arches = _g(cfg, "architectures") or []
+    return any("qwen" in a.lower() for a in arches)
+
+
+def parse(cfg: Any) -> ModelIR:
+    text_cfg  = _text_config(cfg)
+    arch_name = architecture_name(cfg, "qwen")
+    model_type = (_g(text_cfg, "model_type") or "").lower()
+    is_moe = "moe" in model_type or bool(_g(text_cfg, "num_experts"))
+
+    num_layers  = _g(text_cfg, "num_hidden_layers", 0)
+    num_heads   = _g(text_cfg, "num_attention_heads", 0)
+    num_kv_heads = _g(text_cfg, "num_key_value_heads", num_heads)
+    hidden_size  = _g(text_cfg, "hidden_size", 0)
+    head_dim     = _g(text_cfg, "head_dim") or (hidden_size // num_heads if num_heads else None)
+    activation   = (_g(text_cfg, "hidden_act") or "silu").lower()
+
+    # Standard (non-hybrid) attention kind
+    if num_kv_heads == num_heads:
+        full_attn_kind = "mha"
+    elif num_kv_heads == 1:
+        full_attn_kind = "mqa"
+    else:
+        full_attn_kind = "gqa"
+
+    sliding_window  = _g(text_cfg, "sliding_window")
+    sliding_pattern = _g(text_cfg, "sliding_window_pattern")
+    layer_types     = _g(text_cfg, "layer_types") or []
+
+    # Hybrid linear-attention fields (Qwen3.5/3.6)
+    linear_num_kv_heads = _g(text_cfg, "linear_num_key_heads") or 0
+    linear_head_dim     = _g(text_cfg, "linear_key_head_dim") or head_dim
+
+    # MoE FFN fields
+    num_experts          = _g(text_cfg, "num_experts") or 0
+    num_experts_per_tok  = _g(text_cfg, "num_experts_per_tok") or _g(text_cfg, "top_k") or 0
+    num_shared_experts   = _g(text_cfg, "num_shared_experts") or 0
+    moe_intermediate_size = _g(text_cfg, "moe_intermediate_size") or 0
+    dense_intermediate_size = _g(text_cfg, "intermediate_size") or 0
+
+    layers = []
+    for i in range(num_layers):
+        layer_type = layer_types[i] if i < len(layer_types) else "full_attention"
+        is_linear = layer_type == "linear_attention"
+
+        if is_linear:
+            # SSM / recurrent linear-attention layer — no positional mask, compact heads
+            attn = AttentionSpec(
+                kind="linear",
+                num_heads=num_heads,
+                num_kv_heads=linear_num_kv_heads or num_kv_heads,
+                head_dim=linear_head_dim,
+                mask="causal",
+            )
+        else:
+            if sliding_pattern and sliding_window:
+                mask = "sliding" if (i % sliding_pattern) != (sliding_pattern - 1) else "causal"
+                win  = sliding_window if mask == "sliding" else None
+            elif sliding_window:
+                mask, win = "sliding", sliding_window
+            else:
+                mask, win = "causal", None
+
+            attn = AttentionSpec(
+                kind=full_attn_kind,
+                num_heads=num_heads,
+                num_kv_heads=num_kv_heads,
+                head_dim=head_dim,
+                mask=mask,
+                window_size=win,
+            )
+
+        if is_moe and num_experts:
+            ffn = FFNSpec(
+                kind="moe",
+                activation=activation,
+                intermediate_size=dense_intermediate_size or moe_intermediate_size,
+                gated=True,
+                num_experts=num_experts,
+                num_experts_per_tok=num_experts_per_tok,
+                num_shared_experts=num_shared_experts,
+                expert_intermediate_size=moe_intermediate_size,
+            )
+        else:
+            ffn = FFNSpec(
+                kind="dense",
+                activation=activation,
+                intermediate_size=dense_intermediate_size,
+                gated=True,
+            )
+
+        layers.append(decoder_layer(i, attn, ffn, hidden_size))
+
+    vocab_size = _g(text_cfg, "vocab_size", 0) or _g(cfg, "vocab_size", 0)
+    tie_word_embeddings = bool(
+        _g(text_cfg, "tie_word_embeddings", _g(cfg, "tie_word_embeddings", False))
+    )
+    return ModelIR(
+        name=model_name(cfg, arch_name),
+        architecture=arch_name,
+        vocab_size=vocab_size,
+        hidden_size=hidden_size,
+        max_position_embeddings=_g(text_cfg, "max_position_embeddings"),
+        tie_word_embeddings=tie_word_embeddings,
+        layers=layers,
+        extras=decoder_extras(vocab_size, hidden_size, tie_word_embeddings),
+    )
+
+
+def _text_config(cfg: Any) -> Any:
+    model_type = (_g(cfg, "model_type") or "").lower()
+    if model_type in _WRAPPED_TYPES:
+        sub = _g(cfg, "text_config")
+        if sub is not None:
+            return sub
+    return cfg

{model_unfolder-0.2.3 → model_unfolder-0.2.4}/model_unfolder/diagram.py

@@ -44,6 +44,11 @@ class Diagram:
         """Return parameter-count estimates: total / active / per-layer breakdown."""
         return self._params
 
+    @property
+    def warnings(self) -> list[str]:
+        """Adapter-emitted warnings — unknown model types, unrecognised layer types, etc."""
+        return list(self.ir.warnings)
+
     def _repr_html_(self) -> str:
         """Jupyter calls this; returned HTML string is rendered inline."""
         return self._html(standalone=False)
@@ -87,9 +92,12 @@ class Diagram:
         return self._html_cache[standalone]
 
     def __repr__(self) -> str:
-        return (
+        s = (
             f"<Diagram {self.ir.name!r} · {self.ir.num_layers} layers · "
             f"~{humanize(self._params['total'])} params"
             + (f" ({humanize(self._params['active'])} active)" if self._params['is_sparse'] else "")
             + ">"
         )
+        if self.ir.warnings:
+            s += "\n" + "\n".join(f"  ⚠ {w}" for w in self.ir.warnings)
+        return s

{model_unfolder-0.2.3 → model_unfolder-0.2.4}/model_unfolder/ir.py

@@ -81,6 +81,7 @@ class ModelIR:
     layers: list                    # list[LayerSpec]
     cross_layer_edges: list = field(default_factory=list)
     extras: dict = field(default_factory=dict)
+    warnings: list = field(default_factory=list)  # adapter-emitted gaps / unknowns
 
     def to_dict(self) -> dict:
         # Avoid dataclasses.asdict here: it recursively deepcopy()s every
@@ -97,6 +98,7 @@ class ModelIR:
             "layers": [_layer_to_dict(layer) for layer in self.layers],
             "cross_layer_edges": [_cross_edge_to_dict(edge) for edge in self.cross_layer_edges],
             "extras": self.extras,
+            "warnings": self.warnings,
         }
 
     @property

{model_unfolder-0.2.3 → model_unfolder-0.2.4}/model_unfolder.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: model-unfolder
-Version: 0.2.3
+Version: 0.2.4
 Summary: Unfold any HuggingFace transformer into an interactive architecture diagram, inline in Jupyter.
 Author: model-unfolder contributors
 License: Apache-2.0
@@ -110,7 +110,8 @@ Param estimates are close to published numbers — DeepSeek-V3 reports `~675B (~
 |---|---|
 | DeepSeek | DeepSeek-V2, DeepSeek-V3, Kimi K2 |
 | Llama | Llama 3 / 3.1 / 3.2 / 3.3 |
-| Mistral | v0.1 / 0.2 / 0.3 | 
+| Mistral | Mistral 7B, Mixtral 8x7B / 8x22B, Mistral Medium 3.5 |
+| Qwen | Qwen2 / 2.5, Qwen2-MoE, Qwen3, Qwen3-MoE, Qwen3.5 / 3.6 (hybrid linear+full attn) |
 | Gemma | Gemma 3, Gemma 4 (31B, E2B, E4B) |
 
 ### Diffusors

{model_unfolder-0.2.3 → model_unfolder-0.2.4}/model_unfolder.egg-info/SOURCES.txt

@@ -22,8 +22,11 @@ model_unfolder/adapters/transformer/blocks.py
 model_unfolder/adapters/transformer/common.py
 model_unfolder/adapters/transformer/families/__init__.py
 model_unfolder/adapters/transformer/families/deepseek.py
+model_unfolder/adapters/transformer/families/fallback.py
 model_unfolder/adapters/transformer/families/gemma4.py
 model_unfolder/adapters/transformer/families/llama.py
+model_unfolder/adapters/transformer/families/mistral.py
+model_unfolder/adapters/transformer/families/qwen.py
 model_unfolder/adapters/transformer/special_parts/__init__.py
 model_unfolder/adapters/transformer/special_parts/per_layer_embedding.py
 model_unfolder/renderers/__init__.py

{model_unfolder-0.2.3 → model_unfolder-0.2.4}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "model-unfolder"
-version = "0.2.3"
+version = "0.2.4"
 description = "Unfold any HuggingFace transformer into an interactive architecture diagram, inline in Jupyter."
 readme = "README.md"
 requires-python = ">=3.9"

model_unfolder-0.2.3/model_unfolder/adapters/transformer/families/__init__.py

@@ -1,12 +0,0 @@
-"""Transformer model-family adapters.
-
-These modules translate family-specific HuggingFace config dialects into the
-shared transformer IR pieces in ``model_unfolder.adapters.transformer``.
-"""
-
-from . import deepseek, gemma4, llama
-
-# Order matters: more specific adapters first.  ``gemma4`` claims its own
-# top-level ``model_type`` / architecture and must run before the generic
-# llama-family matcher.
-ADAPTERS = [deepseek, gemma4, llama]