Add glm 4.7 flash missing docs

Makefile

@@ -15,7 +15,6 @@ help:
 	@echo " make test-saving             Model save/load"
 	@echo " make test-data-format        channels_first/last"
 	@echo " make test-data-format-gpu    channels_first on TF GPU"
-	@echo " make test-links              Link validation (slow)"
 	@echo " make test-gpu                All GPU-only tests"
 	@echo ""
 

README.md

@@ -87,6 +87,7 @@ Documentation sources are also available in [`docs/`](docs/).
     | GLM-4 (GLM-4-9B) | [ChatGLM: A Family of Large Language Models from GLM-130B to GLM-4](https://arxiv.org/abs/2406.12793) | `transformers` |
     | GLM-4-0414 | [THUDM/GLM-4-9B-0414](https://huggingface.co/THUDM/GLM-4-9B-0414) | `transformers` |
     | GLM-4.5 / GLM-4.6 (MoE) | [GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models](https://arxiv.org/abs/2508.06471) | `transformers` |
+    | GLM-4.7-Flash (MoE Lite) | [zai-org/GLM-4.7-Flash](https://huggingface.co/zai-org/GLM-4.7-Flash) | `transformers` |
     | GLM-5 / GLM-5.1 / GLM-5.2 (MoE) | [GLM-5 Technical Report](https://arxiv.org/abs/2602.15763) | `transformers` |
 
 <br>

docs/glm4_moe_lite.md

@@ -0,0 +1,144 @@
+# GLM-4.7-Flash (GLM-4 MoE Lite)
+
+<div class="kf-note kf-note--weights">
+<b>Weights:</b> preconverted Keras weights are hosted at
+<a href="https://huggingface.co/zeromodels/glm-4.7-flash">zeromodels/glm-4.7-flash</a>.
+Load the model and tokenizer with
+<code>from_weights("zeromodels/glm-4.7-flash")</code>. See
+<a href="../loading_weights/">Loading Weights</a>.
+</div>
+
+GLM-4.7-Flash is a compact Mixture-of-Experts text model implemented in pure
+Keras 3. Its ZeroModels architecture name is `glm4_moe_lite`. The decoder uses
+DeepSeek-V3-style Multi-head Latent Attention (MLA) and aux-loss-free
+DeepSeekMoE routing. Unlike GLM-5, it has no DSA sparse-attention indexer.
+
+All experts remain resident in memory even though only four routed experts are
+active for each token.
+
+Links:
+
+- Model card: [zai-org/GLM-4.7-Flash](https://huggingface.co/zai-org/GLM-4.7-Flash)
+- Keras weights: [zeromodels/glm-4.7-flash](https://huggingface.co/zeromodels/glm-4.7-flash)
+- Related architecture: [GLM-4.5 (GLM-4 MoE)](glm4_moe.md)
+- Related architecture: [GLM-5 MoE](glm5_moe.md)
+
+## Variant
+
+GLM-4 MoE Lite currently has one hosted variant.
+
+| Variant | Hosted | Upstream |
+|---|---|---|
+| `glm-4.7-flash` | `zeromodels/glm-4.7-flash` | [`zai-org/GLM-4.7-Flash`](https://huggingface.co/zai-org/GLM-4.7-Flash) |
+
+## API
+
+The family exports these public classes:
+
+- `Glm4MoeLiteConfig`
+- `Glm4MoeLiteModel`
+- `Glm4MoeLiteTextGenerate`
+- `Glm4MoeLiteTokenizer`
+
+### `Glm4MoeLiteModel`
+
+The decoder backbone returns
+`{"last_hidden_state": (batch, sequence, embed_dim)}`.
+
+| Arg | Default | Meaning |
+|---|---|---|
+| `vocab_size` | `154880` | token vocabulary size |
+| `embed_dim` | `2048` | model width |
+| `num_layers` | `47` | decoder blocks |
+| `num_heads` | `20` | query heads |
+| `mlp_dim` | `10240` | dense-layer SwiGLU width |
+| `moe_mlp_dim` | `1536` | per-expert SwiGLU width |
+| `num_experts` | `64` | routed expert count |
+| `num_experts_per_tok` | `4` | routed experts selected per token |
+| `n_shared_experts` | `1` | shared expert count |
+| `n_group` | `1` | expert routing group count |
+| `topk_group` | `1` | routing groups retained per token |
+| `norm_topk_prob` | `True` | renormalize selected routing weights |
+| `routed_scaling_factor` | `1.8` | selected-expert output scale |
+| `first_k_dense` | `1` | leading dense decoder layers |
+| `q_lora_rank` | `768` | query projection bottleneck |
+| `kv_lora_rank` | `512` | key/value projection bottleneck |
+| `qk_nope_head_dim` | `192` | non-rotary query/key width per head |
+| `qk_rope_head_dim` | `64` | rotary query/key width per head |
+| `v_head_dim` | `256` | value width per head |
+| `rope_theta` | `1000000.0` | rotary base frequency |
+| `rope_scaling` | `None` | optional Hugging Face YaRN scaling dictionary |
+| `norm_eps` | `1e-5` | RMSNorm epsilon |
+| `max_position_embeddings` | `202752` | maximum configured context length |
+| `tie_embeddings` | `False` | reuse token embeddings for the LM head |
+
+### `Glm4MoeLiteTextGenerate`
+
+`Glm4MoeLiteModel` plus the language-model head. It returns
+`{"logits": (batch, sequence, vocab_size)}` and adds `.generate()`.
+
+```python
+generate(
+    input_ids,
+    attention_mask=None,
+    max_new_tokens=None,
+    eos_token_id=None,
+    sampler=None,
+    seed=None,
+    **prefill_inputs,
+)
+```
+
+When `eos_token_id` is omitted, the generation class recognizes the model's
+three end markers: `154820`, `154827`, and `154829`.
+
+### `Glm4MoeLiteTokenizer`
+
+The BPE tokenizer uses the `tokenizers` backend and loads `tokenizer.json` from
+the Hub repository unless a local file is provided.
+
+```python
+Glm4MoeLiteTokenizer(hf_id=None, tokenizer_file=None)
+```
+
+Calling it with a string or list of strings returns token IDs. Decode one
+sequence with `.decode(ids)` or a batch with `.batch_decode(ids)`.
+
+## End-to-end example
+
+```python
+import os
+
+os.environ["KERAS_BACKEND"] = "torch"  # or "jax" / "tensorflow"
+
+from zeromodels.models.glm4_moe_lite import (
+    Glm4MoeLiteTextGenerate,
+    Glm4MoeLiteTokenizer,
+)
+
+weights = "zeromodels/glm-4.7-flash"
+model = Glm4MoeLiteTextGenerate.from_weights(
+    weights,
+    load_dtype="bfloat16",
+)
+tokenizer = Glm4MoeLiteTokenizer.from_weights(weights)
+
+inputs = tokenizer("Explain mixture-of-experts routing in one sentence.")
+outputs = model.generate(**inputs, max_new_tokens=64)
+print(tokenizer.decode(outputs[0]))
+```
+
+The upstream model is large despite its sparse activation. Make sure the
+selected backend has enough host and accelerator memory before converting it.
+
+## Backbone example
+
+```python
+from zeromodels.models.glm4_moe_lite import Glm4MoeLiteModel
+
+backbone = Glm4MoeLiteModel.from_weights(
+    "zeromodels/glm-4.7-flash",
+    load_dtype="bfloat16",
+)
+hidden = backbone(inputs)["last_hidden_state"]
+```

tests/integration/test_auto_registry.py

@@ -9,6 +9,7 @@
 """
 
 import inspect
+from pathlib import Path
 
 import pytest
 
@@ -418,3 +419,84 @@ def family(class_name):
         "would return a mismatched (model, config, processor) triple; fix the mapping "
         f"(or allowlist a genuinely-shared type with its exact family set): {bad}"
     )
+
+
+# Families intentionally covered by a shared or differently named documentation page.
+# Keep this mapping explicit: any new package must get its own docs page or be added here
+# with the page that documents it.
+_SHARED_MODEL_DOCS = {
+    "classification_backbones.md": {
+        "cait",
+        "convmixer",
+        "convnext",
+        "convnextv2",
+        "deit",
+        "densenet",
+        "efficientformer",
+        "efficientnet",
+        "efficientnet_lite",
+        "efficientnetv2",
+        "flexivit",
+        "inception_next",
+        "inception_resnetv2",
+        "inceptionv3",
+        "inceptionv4",
+        "levit",
+        "maxvit",
+        "mit",
+        "mlp_mixer",
+        "mobilenetv2",
+        "mobilenetv3",
+        "mobilenetv4",
+        "nextvit",
+        "pit",
+        "poolformer",
+        "regnet",
+        "res2net",
+        "resmlp",
+        "resnet",
+        "resnetv2",
+        "resnext",
+        "senet",
+        "swin",
+        "swinv2",
+        "vgg",
+        "vit",
+        "xception",
+    },
+    "deberta.md": {"deberta_v2", "deberta_v3"},
+    "dinov2.md": {"dino_v2"},
+    "dinov3.md": {"dino_v3"},
+}
+
+
+def test_every_model_family_is_documented():
+    """Every model package has its own page or an explicit shared-page entry."""
+    root = Path(__file__).resolve().parents[2]
+    docs_dir = root / "docs"
+    model_dirs = {
+        path.name
+        for path in (root / "zeromodels" / "models").iterdir()
+        if path.is_dir() and not path.name.startswith("_")
+    }
+    own_pages = {path.stem for path in docs_dir.glob("*.md")}
+
+    shared_families = set()
+    for page, families in _SHARED_MODEL_DOCS.items():
+        assert (docs_dir / page).is_file(), (
+            f"shared model docs page does not exist: {page}"
+        )
+        overlap = shared_families & families
+        assert not overlap, (
+            f"model families listed under multiple shared pages: {overlap}"
+        )
+        shared_families.update(families)
+
+    stale = sorted(shared_families - model_dirs)
+    assert not stale, f"shared model docs entries without a model package: {stale}"
+
+    missing = sorted(model_dirs - own_pages - shared_families)
+    assert not missing, (
+        "model family package(s) without documentation; add docs/<family>.md or an "
+        f"explicit _SHARED_MODEL_DOCS entry: {missing}"
+    )

website/mkdocs.yml

@@ -145,6 +145,7 @@ nav:
           - GLM: glm.md
           - GLM-4: glm4.md
           - GLM-4 MoE: glm4_moe.md
+          - GLM-4 MoE Lite: glm4_moe_lite.md
           - GLM-5 MoE: glm5_moe.md
           - GPT: gpt.md
           - GPT-2: gpt2.md