Update README.md

README.md

@@ -7,16 +7,8 @@ tags:
   - text-to-image
   - image-to-image
 ---
+# NLD-Uni
 
-# Nemotron-Labs-Diffusion-Uni-8B
-
-## Description
-
-NLD-Uni is a unified multimodal model for image understanding and image generation. It combines a diffusion language model with specialized visual components and a mixture-of-transformers design. Understanding and generation use distinct transformer parameter paths within a common multimodal framework.
-
-For image understanding, NLD-Uni encodes images into continuous visual features and generates text responses. For image generation, it models discrete visual tokens through iterative diffusion and decodes those tokens into RGB images. The architecture supports autoregressive and blockwise diffusion decoding for text, as well as image-conditioned generation and editing when enabled by the checkpoint and inference implementation.
-
-This card describes the NLD-Uni architecture generally.
 
 ## Environment
 
@@ -53,6 +45,23 @@ print(answer)
 
 Image inputs accept Pillow images or local file paths. Image methods return Pillow images; `generate` returns text by default. Token IDs can be supplied through `input_ids` for text prompts, and `return_token_ids=True` returns the generated token tensor. The `inference.py` command exposes each task.
 
+## Default inference recipes
+
+| Parameter | Text-to-image | Image editing | Understanding |
+|---|---|---|---|
+| Resolution | 1024 × 1024 | 1024 × 1024 internally | Native aspect ratio with RADIO resizing |
+| Activation steps | 64 | 64 | 32 steps per 32-token block |
+| Semantic refinement stages | 4 | 4 | — |
+| Guidance | 3.0 | Text 1.6, image 2.0, edit mode 1 | — |
+| Temperature | 0.8 | 0.8 | 0.0 |
+| Confidence policy | Stratified | Stratified | Threshold 0.9 |
+| Schedule | Shift 5 | Shift 5 | Block diffusion |
+| Maximum new tokens | — | — | 16000 |
+
+Image sampling uses dynamic ranking temperature (5.0 down to 0.5, cosine-squared schedule), full-attention approximation, prompt caching, and checkpoint-native logit alignment. Four refinement stages add three trailing model calls to the 64 activation rounds. `refinement_steps` accepts 1, 4, 8, or 16; 1 is the direct root-to-token inference control.
+
+The public image methods expose `resolution`, `steps`, `refinement_steps`, `guidance_scale`, `temperature`, `seed`, `top_k`, `top_p`, `shift`, and `show_progress`. Editing additionally exposes `image_guidance_scale` and `crop_to_source`; source-padding VQ codes are preserved. The default crops away square padding and restores the source dimensions. Understanding exposes `max_new_tokens`, `block_length`, `steps_per_block`, `threshold`, and `temperature`. No thinking prefix is added. Rotary phase calculations stay in FP32 even inside an external autocast context.
+
 ## Minimal command-line inference
 
 ```bash
@@ -62,168 +71,3 @@ python inference.py /path/to/hf_checkpoint --task understand --image blue.png --
 ```
 
 Replace the model argument with the published Hugging Face model ID for remote loading. The standalone script only uses Transformers and this repository's model code.
-
-
-### License/Terms of Use
-
-GOVERNING TERMS: Use of this model is governed by the [NVIDIA Open Source License Agreement](https://huggingface.co/nvidia/NL-Diffusion-Image/raw/main/LICENSE).
-
-## Deployment Geography
-
-Global, subject to the governing release terms.
-
-## Use Case
-
-NLD-Uni is intended for research and evaluation of unified multimodal systems, including:
-
-- Visual question answering and image-grounded reasoning.
-- Document, chart, and scene-text understanding.
-- Visual grounding and interface localization using task-specific prompts and output formats.
-- Text-to-image generation.
-- Image-conditioned generation and editing for checkpoints trained for these tasks.
-
-## Release Date
-
-10/27/2026
-
-## Model Architecture
-
-**Architecture Type:** Transformer
-
-**Network Architecture:** Diffusion language model with understanding and generation transformer paths, a continuous visual encoder for understanding, and a discrete image tokenizer for generation.
-
-**Language-model foundation:** Nemotron diffusion language model.
-
-## Input
-
-**Input Type(s):** Text and images.
-
-**Input Format(s):** Text strings and image objects or tensors accepted by the supplied processor.
-
-**Input Parameters:** One-dimensional text-token sequences and two-dimensional image content, represented internally as visual features or discrete image tokens according to the task.
-
-**Other Properties Related to Input:** Context length, image resolution, visual-token budget, and preprocessing are checkpoint- and task-dependent. The released inference configuration defines the supported limits; a universal 900-token prompt limit is not specified for the architecture.
-
-## Output
-
-**Output Type(s):**
-
-- Text
-- Image
-
-**Output Format(s):**
-
-- **Text:** `.json` (UTF-8 JSON records containing generated text).
-- **Image:** `.png` (RGB images).
-
-These serialized formats are implemented by `validate.py`. The inference API returns text strings or token IDs for understanding and decoded image objects or tensors for generation. Grounding outputs follow the requested coordinate or location-token format.
-
-**Output Parameters:** One-dimensional token sequences or images with height, width, and three color channels. Tensor layout and pixel range are defined by the inference API.
-
-**Other Properties Related to Output:** Text output length and generated image resolution depend on the checkpoint and inference settings. Configurations for 1,024 × 1,024 image generation are implemented; the release should list its validated resolutions.
-
-## Software Integration
-
-NLD-Uni’s native implementation uses PyTorch with CUDA-accelerated execution. The codebase also includes a dedicated SGLang integration for the image-understanding and text-generation path. This does not establish SGLang support for the image-generation path.
-
-**Runtime Engine(s):**
-
-- PyTorch for native multimodal inference.
-- SGLang with the NLD-Uni/LaVida RADIO-H adapter for supported understanding configurations.
-
-**Reference Platform:** Linux with NVIDIA CUDA GPUs.
-
-**Reference Evaluation Hardware:** NVIDIA H100 (Hopper).
-
-## Model Version(s)
-
-- **NLD-Uni:** Unified image-understanding and image-generation architecture based on LaVida-O.
-
-Training stage, checkpoint revision, tokenizer revision, and inference configuration are part of the identity of an evaluated model. Results from earlier LaVida-O or image-only diffusion models should retain their original model names.
-
-# Training and Evaluation Datasets
-
-## Training Dataset
-
-More than 100M image-text pairs.
-
-**Data Collection Method by dataset**
-
-- Hybrid: Automated, Manually-Collected, Synthetic
-
-**Labeling Method by dataset**
-
-- Hybrid: Automated, Manually-Labeled, Synthetic
-
-## Evaluation Datasets
-
-### Documented Functional Validation Dataset
-
-- **Modalities:** Text and RGB images.
-- **Nature of content:** Synthetic everyday-object, animal, and landscape scenes; text-to-image prompts, edit instructions, and image-understanding questions. The generated images are used as inputs to subsequent editing and understanding tasks.
-- **Linguistic characteristics:** English prompts, edit instructions, questions, and recorded answers.
-- **Collection and labeling:** Three scripted scene cases in `validate.py`; model-generated images and answers were manually inspected for prompt alignment and editing fidelity.
-
-Counts and results are recorded in [VALIDATION.md](VALIDATION.md), [validation.json](validation.json), and [examples/results.json](examples/results.json). These are functional smoke tests, not measurements of benchmark accuracy.
-
-### Benchmark Evaluation Properties
-
-The benchmark suite below covers image and text inputs, including mathematical and scientific questions, documents, charts, scene text, interfaces, referring expressions, and image-generation prompts. OCRBenchV2-EN identifies an English subset; the exact language coverage of the other evaluated splits must be documented from the evaluation records.
-
-### Image Understanding
-
-The evaluation suite should identify the following benchmarks and splits explicitly:
-
-| Benchmark | Evaluation target | Primary reporting requirement |
-|---|---|---|
-| MathVista_MINI | Visual mathematical reasoning | Full 1,000-example accuracy; identify answer-extraction judge |
-| MMMU_DEV_VAL | Multidisciplinary visual reasoning | Report development and validation splits separately |
-| AI2D_TEST | Diagram understanding | Accuracy and image/prompt protocol |
-| ChartQA_TEST | Chart question answering | Relaxed accuracy and subset aggregation |
-| DocVQA_VAL | Document question answering | ANLS |
-| TextVQA_VAL | Scene-text question answering | VQA accuracy with specified normalization |
-| RealWorldQA | Real-world visual understanding | Accuracy |
-| InfoVQA_VAL | Infographic question answering | ANLS |
-| ScienceQA_TEST | Science question answering | Identify the image-containing subset when used |
-| ScreenSpot-Pro | Interface grounding | Point-localization accuracy; separate first-pass and crop-refined results |
-| RefCOCO | Referring-expression grounding | IoU ≥ 0.5 accuracy for val, testA, and testB; disclose the native LOC protocol |
-| OCRBenchV2-EN | OCR and document understanding | Official English category aggregation |
-
-### Image Generation
-
-| Benchmark | Evaluation target | Primary metric |
-|---|---|---|
-| GenEval | Object presence, counts, colors, positions, and attribute binding | Overall and category scores |
-| DPG-Bench | Alignment with detailed compositional prompts | Overall and category scores |
-| MJHQ-30K | Generated-image distribution and quality | FID; any additional preference metric must identify its model and version |
-
-Evaluation should identify the exact checkpoint, prompt processing, image preprocessing, decoding parameters, seed policy, sample count, and evaluator version. Report reasoning and non-reasoning results separately. Tokenizer reconstruction evaluation, if performed, should be identified separately from end-to-end image-generation quality.
-
-## Key Considerations
-
-NLD-Uni can produce incorrect text, inaccurate grounding coordinates, and synthetic images that do not fully follow the prompt. Generated content may reflect biases in training data or contain inappropriate material. Image-conditioned outputs may change details that users intended to preserve.
-
-Applications should be evaluated on their intended tasks and data, with appropriate input and output controls. Model outputs should not be treated as verified factual evidence solely because they are fluent or visually plausible.
-
-## Inference
-
-**Native Engine:** PyTorch.
-
-**Understanding Serving Integration:** SGLang with the model-specific adapter, for validated configurations.
-
-**Reference Precision:** BF16 where supported by the selected implementation.
-
-**Reference Hardware:** NVIDIA H100.
-
-Use the model-specific tokenizer, processor, chat template, and task interface. Image understanding uses the continuous visual encoder; image generation uses the discrete image-token pipeline. Runtime optimizations, including CUDA graphs, should be evaluated with the exact checkpoint and decode configuration before publishing accuracy or speed claims.
-
-## Ethical Considerations
-
-NVIDIA believes trustworthy AI is a shared responsibility. Developers should evaluate NLD-Uni for the intended application, including potential errors, bias, privacy concerns, and misuse of generated content.
-
-Users should have appropriate rights and permissions for input images and other supplied content. Generated or edited images may alter identities, text, object proportions, or scene details. Applications should communicate the synthetic nature of generated content where appropriate.
-
-Users are responsible for model inputs and outputs and for implementing appropriate safeguards before deployment.
-
-Report model quality concerns, security vulnerabilities, or NVIDIA AI concerns through [NVIDIA’s reporting channel](https://www.nvidia.com/en-us/support/submit-security-vulnerability/).
-

attention.py

@@ -75,9 +75,9 @@ class MinistralFlexAttention(Ministral3Attention):
         query_states = query_states * _get_llama_4_attn_scale(cache_position, self.config.rope_parameters.get('llama_4_scaling_beta'), self.config.rope_parameters.get('original_max_position_embeddings')).to(query_states.dtype)
         current_key_states = key_states
         current_value_states = value_states
-        cached_routing_metadata = getattr(past_key_values, '_nlduni_dual_tower_routing', None)
+        cached_routing_metadata = getattr(past_key_values, '_lavida_dual_tower_routing', None)
         isolate_generation_flash = overwrite_attn_impl == 'flash_attn' and self.split_tower and (routing is not None or cached_routing_metadata is not None)
-        cached_has_generation = bool(getattr(past_key_values, '_nlduni_dual_tower_has_generation', False))
+        cached_has_generation = bool(getattr(past_key_values, '_lavida_dual_tower_has_generation', False))
         drop_cached_understanding = isolate_generation_flash and routing is True and (not cached_has_generation)
         if past_key_values is not None:
             if use_cache:

chat_utils.py

@@ -17,7 +17,7 @@ class SimpleOutputWithPast(ModelOutput):
     hidden_states: tuple[torch.FloatTensor, ...] | None = None
     attentions: tuple[torch.FloatTensor, ...] | None = None
 from .utils import maybe_truncate_last_dim, pad_along_last_dim
-from .configuration_nlduni import get_ministral_shift_sbd_logits
+from .configuration_lavida import get_ministral_shift_sbd_logits
 from .sbd_attention import build_image_attention_doc_ids
 
 def _resolve_shift_sbd_logits(model, shift_logits):

config.json

@@ -23,14 +23,14 @@
   "eos_token_id": 11,
   "pad_token_id": 11,
   "tie_word_embeddings": false,
-  "model_type": "nlduni",
+  "model_type": "lavida_semantic_tree",
   "architectures": [
-    "NLDUniForCausalLM"
+    "LaVidaForCausalLM"
   ],
   "torch_dtype": "bfloat16",
   "auto_map": {
-    "AutoConfig": "configuration_nlduni.NLDUniConfig",
-    "AutoModel": "modeling_nlduni.NLDUniForCausalLM",
-    "AutoModelForCausalLM": "modeling_nlduni.NLDUniForCausalLM"
+    "AutoConfig": "configuration_lavida.LaVidaConfig",
+    "AutoModel": "modeling_lavida.LaVidaForCausalLM",
+    "AutoModelForCausalLM": "modeling_lavida.LaVidaForCausalLM"
   }
 }

configuration_nlduni.py → configuration_lavida.py

@@ -1,7 +1,7 @@
 from transformers import PretrainedConfig
 
-class NLDUniConfig(PretrainedConfig):
-    model_type = 'nlduni'
+class LaVidaConfig(PretrainedConfig):
+    model_type = 'lavida_semantic_tree'
     # Architecture invariants; deliberately absent from serialized configuration.
     d_model_gen = 4096
     d_model = 4096
@@ -53,7 +53,7 @@ class NLDUniConfig(PretrainedConfig):
         values.pop('_name_or_path', None)
         return values
 
-MinistralDLMConfig = NLDUniConfig
+MinistralDLMConfig = LaVidaConfig
 
 def get_ministral_shift_sbd_logits(config):
     return True

core.py

@@ -17,7 +17,7 @@ from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_u
 from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from transformers.processing_utils import Unpack
 from transformers.utils import TransformersKwargs, auto_docstring
-from .configuration_nlduni import MinistralDLMConfig
+from .configuration_lavida import MinistralDLMConfig
 
 def rotate_half(x):
     x1 = x[..., :x.shape[-1] // 2]
@@ -387,7 +387,7 @@ class Ministral3Model(Ministral3PreTrainedModel):
         if use_cache and past_key_values is None:
             past_key_values = DynamicCache()
         past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None and (partial_dual_tower or cache_position is None) else 0
-        stored_cached_routing = getattr(past_key_values, '_nlduni_dual_tower_routing', None)
+        stored_cached_routing = getattr(past_key_values, '_lavida_dual_tower_routing', None)
         track_dual_tower_routing = partial_dual_tower and (canonical_routing is not None or stored_cached_routing is not None)
         current_routing = _routing_mask(canonical_routing, inputs_embeds) if track_dual_tower_routing else None
         cached_routing = stored_cached_routing
@@ -398,8 +398,8 @@ class Ministral3Model(Ministral3PreTrainedModel):
         elif track_dual_tower_routing and cached_routing.device != current_routing.device:
             raise ValueError('cached dual-tower routing and inputs must share a device')
         if track_dual_tower_routing and past_key_values is not None:
-            past_key_values._nlduni_dual_tower_routing = cached_routing
-            past_key_values._nlduni_dual_tower_has_generation = bool(getattr(past_key_values, '_nlduni_dual_tower_has_generation', False))
+            past_key_values._lavida_dual_tower_routing = cached_routing
+            past_key_values._lavida_dual_tower_has_generation = bool(getattr(past_key_values, '_lavida_dual_tower_has_generation', False))
         if cache_position is None:
             cache_position = torch.arange(past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device)
         if position_ids is None:
@@ -420,7 +420,7 @@ class Ministral3Model(Ministral3PreTrainedModel):
             layer_kwargs = split_layer_kwargs if split_layer_kwargs is not None and layer_idx >= self.dual_tower_layers else kwargs
             hidden_states = decoder_layer(hidden_states, attention_mask=causal_mask, position_ids=position_ids, past_key_values=past_key_values, use_cache=use_cache, cache_position=cache_position, position_embeddings=position_embeddings, routing=canonical_routing, **layer_kwargs)
         if track_dual_tower_routing and use_cache and (past_key_values is not None):
-            past_key_values._nlduni_dual_tower_routing = torch.cat((cached_routing, current_routing), dim=1)
-            past_key_values._nlduni_dual_tower_has_generation = past_key_values._nlduni_dual_tower_has_generation or canonical_routing is not None
+            past_key_values._lavida_dual_tower_routing = torch.cat((cached_routing, current_routing), dim=1)
+            past_key_values._lavida_dual_tower_has_generation = past_key_values._lavida_dual_tower_has_generation or canonical_routing is not None
         hidden_states = _routed_module(hidden_states, canonical_routing, self.norm, getattr(self, 'norm_gen', None))
         return BaseModelOutputWithPast(last_hidden_state=hidden_states, past_key_values=past_key_values if use_cache else None)

model-00001-of-00008.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c4d490e010475dde3e2b57de64261a56d6bf8e9ca7d077754913fd628f7de5f6
+oid sha256:3b36eb1336d3c0b9dae232de88c8ceb8c442eba5df92019a3cc788663b90e5c6
 size 4974634832

model-00002-of-00008.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:933c4b2f44f9cd7be30ac7a40efbb3c58485aa9fbd6396245d8e2337faf5c5b3
+oid sha256:20c0142736a64977044ad720cb7f575b82d7a31c907b6b51ecb7ffa1e9628e11
 size 4999819672

model-00003-of-00008.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6893a316ec6655eb3f518be0f3ae3dfc80621922741df7d14f12b3d9fd123bf6
+oid sha256:c44643654fae7c8bb4440f91d6c8ff707411552e53a9b290f42b9fa881b38234
 size 4949454016

model-00004-of-00008.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:858fab295f100b897634b72824d01e01f368dc1d500de03916854d7b2769986c
+oid sha256:b90c3460f3cb1580b6ca663d70ec840a8fac2ae6ed043d9ee4450107fb744c56
 size 4999819744

model-00005-of-00008.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0d31608093622df48c2284fa310170f9a1bfd127d7108bdacbc8106fcf296691
+oid sha256:ce7b416181248e1516e40805f9cfdb6e001c064308250fb92094563ce3d6c507
 size 4999819776

model-00006-of-00008.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4356b837d798b2d3250cd067c15b7f5d56fd19a8b8605bb107fb9d727c1726d0
+oid sha256:34ddb4a6a1f67153e9ed34e23a5542a23c9032f0380c77c2c3b31caf8af5db89
 size 4999819272

model-00007-of-00008.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:003f5a6c4c727e5f560e16ad91ee2264ed138c7190108dfc9b68d1b8c8445545
+oid sha256:01bd1ff8c8a8a687033b6f6c0ac51c959235ec748547b0364a98508e2e444f12
 size 4456023446

model-00008-of-00008.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:299783fae2f540dfd723abbb09f8661b57122210d44b23a4c65f24b254542550
+oid sha256:e163845d238de6fc71a9990a2aaf0dea8514de1891a244d6fae9c96016bca92d
 size 2155913464

modeling_nlduni.py → modeling_lavida.py

@@ -7,7 +7,7 @@ from transformers import AutoTokenizer
 from transformers.generation import GenerationMixin
 from transformers.modeling_outputs import CausalLMOutputWithPast
 from transformers.utils.hub import cached_file
-from .configuration_nlduni import NLDUniConfig
+from .configuration_lavida import LaVidaConfig
 from .core import Ministral3Model, Ministral3PreTrainedModel
 from .attention import MinistralFlexAttention
 from .bridge import PackedLinearImageCodec
@@ -20,9 +20,8 @@ from .chat_utils import generate_with_prefix_cache_block_diff
 from .sampling import text_to_image as sample_image
 from .image_utils import process_images, fix_one_prompt, tokenizer_image_token
 from .conversation import Conversation
-from .tree import BalancedSemanticTree, production_semantic_tree_contract
 
-class NLDUniBackbone(Ministral3Model):
+class LaVidaBackbone(Ministral3Model):
     def __init__(self, config):
         super().__init__(config)
         self.vision_tower=RadioVisionTower()
@@ -66,12 +65,12 @@ class NLDUniBackbone(Ministral3Model):
         states=self.packed_image_codec.expand_fixed_grid(hidden_states.reshape(-1,(h//2)*(w//2),4096),gen_shape)
         return self.gen_predictor(states.flatten(0,1))
 
-class NLDUniForCausalLM(Ministral3PreTrainedModel, GenerationMixin):
-    config_class=NLDUniConfig
+class LaVidaForCausalLM(Ministral3PreTrainedModel, GenerationMixin):
+    config_class=LaVidaConfig
     base_model_prefix=''
     def __init__(self, config):
         super().__init__(config)
-        self.encoder=NLDUniBackbone(config)
+        self.encoder=LaVidaBackbone(config)
         self.diffusion_head=nn.Linear(config.hidden_size,config.vocab_size,bias=False)
         self.mask_token_id=100
         self.vocab_size=config.vocab_size
@@ -84,8 +83,7 @@ class NLDUniForCausalLM(Ministral3PreTrainedModel, GenerationMixin):
         target=model[0] if isinstance(model,tuple) else model
         options={k:kwargs[k] for k in ('revision','cache_dir','local_files_only','token','subfolder') if k in kwargs}
         target._tokenizer=AutoTokenizer.from_pretrained(pretrained_model_name_or_path,**options)
-        tree=cached_file(pretrained_model_name_or_path,'pca_tree.pt',**options)
-        BalancedSemanticTree.from_artifact(production_semantic_tree_contract(tree))
+        tree=cached_file(pretrained_model_name_or_path,'semantic_tree.pt',**options)
         target.config.image_token_tree_path=tree
         target.eval()
         return model
@@ -96,8 +94,8 @@ class NLDUniForCausalLM(Ministral3PreTrainedModel, GenerationMixin):
         super().save_pretrained(destination, **kwargs)
         self.tokenizer.save_pretrained(destination)
         asset = Path(self.config.image_token_tree_path)
-        if asset.resolve() != (destination / 'pca_tree.pt').resolve():
-            shutil.copyfile(asset, destination / 'pca_tree.pt')
+        if asset.resolve() != (destination / 'semantic_tree.pt').resolve():
+            shutil.copyfile(asset, destination / 'semantic_tree.pt')
     @property
     def model(self):return self.encoder
     def get_model(self):return self.encoder
@@ -209,8 +207,7 @@ from .core import rotate_half as _bundle_core
 
 from .refinement import _validate_refinement_steps as _bundle_refinement
 
-from .conversation import Conversation
-from .tree import BalancedSemanticTree, production_semantic_tree_contract as _bundle_conversation
+from .conversation import Conversation as _bundle_conversation
 
 from .constants import NemotronMinistralDLMConstant as _bundle_constants
 

release_manifest.json

@@ -1,34 +1,32 @@
 {
-  "revision": "generation-only-step-10000",
   "sha256": {
-    ".gitattributes": "50360a3ded843d95a2667395b3131f6d499b905d941331a4e84e7f0351769023",
+    ".gitattributes": "137fabec805869007ee08b090309aa47a8ce228d8e5caba83bebc5081efe82b8",
     ".gitignore": "e1900f365321556fa39bfe45527be4e15a1de7fe96b71ca8d294b690a1026a8d",
-    "README.md": "50038fb0e318e290c123382d4f6284ce80699bd60d7fabf0ee6db27419e12b07",
-    "attention.py": "5617bdd53aaa0d6f0e79745c5e811f36335a0c1f458218ab1fc02c4606d79f37",
+    "README.md": "801688e7986cc673c14c65cfd86db420f528a2d189c547230e902cbf7ce0c227",
+    "attention.py": "73ba82226b434fec56dcaeb2268c794cae9a62a4ed2769c65a390e16a28ebc77",
     "bridge.py": "0e20f9ad1c81912489dd7c5063507b7fb72ead7d1895467442277e06ee3bc846",
     "chat_template.jinja": "b80135b85225bef1a5129eb2068d0d5df4975edf4ffb102a50de1fb069c9e78a",
-    "chat_utils.py": "0683087dbadea1ce510d77816eed8ccfba998aa6c64833fd93c4e46c6edbc734",
-    "config.json": "9dd7b2514b477712b9e023dec3f0d3749e8ccd9e27869cb4bf70fe9a8097fff0",
-    "configuration_nlduni.py": "fa6d15657c9fadeb1f2bc142ea33b73c225374d5646565a077b43d6fd3e25f38",
+    "chat_utils.py": "22f57a7003165392eab509763e1d5e069056fb1b732e409c6db8fbfc3c57735e",
+    "config.json": "9f57cd590b2debd189cc802ed5eba420a034b9708e4e275727f9001d04488f6b",
+    "configuration_lavida.py": "705696dc266816cb187cf8c495c18e3f9956ff95c292cfedb2258dfadd9a2f30",
     "constants.py": "1149095b05a6da911aae371d4899edcaba154352dc30efd2d5b72bb8767f785f",
     "conversation.py": "5e11082e6f5201978a0eb99523120d1272a6e0b311bfa75b273fa02d927d1002",
-    "core.py": "4dad9ea004daa31236fb1eb628d2ad1b46d310914114bd93c401fd41161d5c66",
+    "core.py": "f95eee180ec24ca1e0ac2ac026b7f9a28ecc96fb756d5d257a44d37eda66759b",
     "generation_config.json": "35e13707e4e588edaa4d7c12dff1d9ccd49efe0e07c99e3392c42b88f9398faa",
     "image_utils.py": "62755735051898eb909937712e0badcf6614d30405ea51d7d33522f484a319d1",
     "inference.py": "37e0b19b3d003621a620b226d6a9d39725e9450a7a6431d992ceea49cb2b92b2",
     "markers.py": "17f69056ff9db548dc223971021a91444ecaef3257df5b5f30fd2e5230bbadc7",
-    "model-00001-of-00008.safetensors": "c4d490e010475dde3e2b57de64261a56d6bf8e9ca7d077754913fd628f7de5f6",
-    "model-00002-of-00008.safetensors": "933c4b2f44f9cd7be30ac7a40efbb3c58485aa9fbd6396245d8e2337faf5c5b3",
-    "model-00003-of-00008.safetensors": "6893a316ec6655eb3f518be0f3ae3dfc80621922741df7d14f12b3d9fd123bf6",
-    "model-00004-of-00008.safetensors": "858fab295f100b897634b72824d01e01f368dc1d500de03916854d7b2769986c",
-    "model-00005-of-00008.safetensors": "0d31608093622df48c2284fa310170f9a1bfd127d7108bdacbc8106fcf296691",
-    "model-00006-of-00008.safetensors": "4356b837d798b2d3250cd067c15b7f5d56fd19a8b8605bb107fb9d727c1726d0",
-    "model-00007-of-00008.safetensors": "003f5a6c4c727e5f560e16ad91ee2264ed138c7190108dfc9b68d1b8c8445545",
-    "model-00008-of-00008.safetensors": "299783fae2f540dfd723abbb09f8661b57122210d44b23a4c65f24b254542550",
+    "model-00001-of-00008.safetensors": "3b36eb1336d3c0b9dae232de88c8ceb8c442eba5df92019a3cc788663b90e5c6",
+    "model-00002-of-00008.safetensors": "20c0142736a64977044ad720cb7f575b82d7a31c907b6b51ecb7ffa1e9628e11",
+    "model-00003-of-00008.safetensors": "c44643654fae7c8bb4440f91d6c8ff707411552e53a9b290f42b9fa881b38234",
+    "model-00004-of-00008.safetensors": "b90c3460f3cb1580b6ca663d70ec840a8fac2ae6ed043d9ee4450107fb744c56",
+    "model-00005-of-00008.safetensors": "ce7b416181248e1516e40805f9cfdb6e001c064308250fb92094563ce3d6c507",
+    "model-00006-of-00008.safetensors": "34ddb4a6a1f67153e9ed34e23a5542a23c9032f0380c77c2c3b31caf8af5db89",
+    "model-00007-of-00008.safetensors": "01bd1ff8c8a8a687033b6f6c0ac51c959235ec748547b0364a98508e2e444f12",
+    "model-00008-of-00008.safetensors": "e163845d238de6fc71a9990a2aaf0dea8514de1891a244d6fae9c96016bca92d",
     "model.safetensors.index.json": "f2c7c2353c2a2eb70e7f80889c4fca85caceba217f1ccd9aa6647a7c644f6e30",
-    "modeling_nlduni.py": "d7a04c330d1cd14fc638297a61a674cfdc666d8eb5ee24c41625ff1875040021",
+    "modeling_lavida.py": "ebd4e6876a77bc4d117786da484df7bbcec924c1a708c77df3c92d86cea93d15",
     "padding.py": "53ba2cc8b37eec837b4fd0a3932a5f2d55051bbe5796c848921ce95171336dce",
-    "pca_tree.pt": "4b7ee94fbb10fd4146a4ad167b595b037e3f591eaa12ecf3f0b411c82384a714",
     "pmj.py": "33920c0bb48ab99bafb3dde99f0e44eff7e6004596777d2e09ec4d07e039bfa3",
     "positions.py": "3e0ee23d4bd59059690bc9962fa0948c18f52af13c1295d5322309ba6472f0c4",
     "projector.py": "b402ae3a9ca1d7867c139e3180f905411d4932901cd7ce19910865db8709691f",
@@ -39,16 +37,16 @@
     "sampling.py": "6111a83d7d737eab9c702af281f209316e7458a785d0102c4f2bf3fd4e315491",
     "sbd_attention.py": "828f02ae59d430485507b00c76d0b85158dfa10dfea7888519ec80522798a134",
     "schedules.py": "243d3fa6924f11397df422c9e50dc37bd4e85d882c518c3a4a75545a3b1a367e",
+    "semantic_tree.pt": "371bde0b74b7fd9c0c4dc43e095f516c797707dbeaf9594206acfcff4e2b02fb",
     "special_tokens_map.json": "57cc1f9b3ea2a038d1107bdf4f4b3c45b5ed9ce175e7410d70070e07a3f35a76",
     "token_constants.py": "6515bb3f128730edcd02907cf49334ac9cae2ed373611b0a260b3bdf4ff678f1",
     "tokenizer.json": "f9ad3482369ba91711ff939f9eeb761d618b3157fcbfc0715f18e34b918eca97",
     "tokenizer_config.json": "5eb18e9f9e0653added53a9200a8d38fdbc1d0cdb4602f8e354f30391f8cb1f3",
-    "tree.py": "b22d647f4d6def2a8df6f1b336db0b03a9cf0131dccf124f949cfa704e227c81",
-    "tree_metadata.json": "16ccc59386d2d73094ba9b2562feff45b7a171f4f37053f9b5ed756d5c44e3d5",
+    "tree.py": "ecb3d15ece7a5a3a19c47d269f941656a33e0956603559d108736a89955d19cd",
     "utils.py": "6a0036a6451918de85fcc9cc3a82f5c195227bda5df58103156dd9941f6da0bc",
     "vision.py": "375c9cd952135701b71b35581a72c46cc58bb9701392343f9817f9cf1e68bb72",
     "vq.py": "3029623e15687b8fe95594b25410cfa2c0a82eebaa821b6ab380d1dcc394e537",
     "vq_model.py": "0f28d77789671b9c05402de7dca7bfccaa57264336ba778a76f3b59966baced7",
-    "weights_manifest.json": "fe3af69d1ff222d3de89efd6305fcb3d396204b3c208491866d4da1137bea6f9"
+    "weights_manifest.json": "76d0c7664ab34a29d59a0444e6c4160acc13cc02810a1429c0afe21438a10973"
   }
 }

pca_tree.pt → semantic_tree.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4b7ee94fbb10fd4146a4ad167b595b037e3f591eaa12ecf3f0b411c82384a714
-size 1052237
+oid sha256:371bde0b74b7fd9c0c4dc43e095f516c797707dbeaf9594206acfcff4e2b02fb
+size 1050709

tree.py

@@ -10,7 +10,7 @@ from pathlib import Path
 import torch
 import torch.nn.functional as F
 PINNED_ARTIFACT_VERSION = 1
-PINNED_ARTIFACT_ALGORITHM = 'recursive_centered_pca_rank_median_v1'
+PINNED_ARTIFACT_ALGORITHM = 'divisive_exact_balanced_binary_kmeans'
 PINNED_ARTIFACT_METRIC = 'normalized_squared_l2_tree_ultrametric'
 PINNED_PRODUCTION_VOCAB_SIZE = 131072
 PINNED_PRODUCTION_DEPTH = 17
@@ -80,7 +80,7 @@ class SemanticTreeArtifactContract:
             raise ValueError(f'Unsupported semantic-tree algorithm: {self.algorithm!r}')
 
 def production_semantic_tree_contract(path: str | os.PathLike[str]) -> SemanticTreeArtifactContract:
-    return SemanticTreeArtifactContract(path=path, sha256='4b7ee94fbb10fd4146a4ad167b595b037e3f591eaa12ecf3f0b411c82384a714', algorithm=PINNED_ARTIFACT_ALGORITHM)
+    return SemanticTreeArtifactContract(path=path, algorithm=None)
 
 def _sha256_file(path: Path) -> str:
     digest = hashlib.sha256()
@@ -143,7 +143,7 @@ class BalancedSemanticTree:
         if not isinstance(artifact, Mapping):
             raise TypeError('semantic-tree artifact must contain a mapping')
         algorithm = artifact.get('algorithm')
-        allowed_algorithms = (contract.algorithm,) if contract.algorithm is not None else (PINNED_ARTIFACT_ALGORITHM,)
+        allowed_algorithms = (contract.algorithm,) if contract.algorithm is not None else PINNED_ARTIFACT_ALGORITHM
         if algorithm not in allowed_algorithms:
             raise ValueError(f'semantic-tree artifact algorithm mismatch: expected one of {allowed_algorithms!r}, got {algorithm!r}')
         pinned_values = {'version': PINNED_ARTIFACT_VERSION, 'metric': PINNED_ARTIFACT_METRIC, 'vocab_size': expected_vocab_size, 'depth': int(math.log2(expected_vocab_size)), 'radii_monotone': True}

tree_metadata.json

@@ -1,7 +0,0 @@
-{
-  "filename": "pca_tree.pt",
-  "algorithm": "recursive_centered_pca_rank_median_v1",
-  "sha256": "4b7ee94fbb10fd4146a4ad167b595b037e3f591eaa12ecf3f0b411c82384a714",
-  "vocab_size": 131072,
-  "depth": 17
-}

weights_manifest.json

@@ -4,49 +4,49 @@
   "shards": [
     {
       "file": "model-00001-of-00008.safetensors",
-      "sha256": "c4d490e010475dde3e2b57de64261a56d6bf8e9ca7d077754913fd628f7de5f6",
+      "sha256": "3b36eb1336d3c0b9dae232de88c8ceb8c442eba5df92019a3cc788663b90e5c6",
       "bytes": 4974634832,
       "tensors": 84
     },
     {
       "file": "model-00002-of-00008.safetensors",
-      "sha256": "933c4b2f44f9cd7be30ac7a40efbb3c58485aa9fbd6396245d8e2337faf5c5b3",
+      "sha256": "20c0142736a64977044ad720cb7f575b82d7a31c907b6b51ecb7ffa1e9628e11",
       "bytes": 4999819672,
       "tensors": 106
     },
     {
       "file": "model-00003-of-00008.safetensors",
-      "sha256": "6893a316ec6655eb3f518be0f3ae3dfc80621922741df7d14f12b3d9fd123bf6",
+      "sha256": "c44643654fae7c8bb4440f91d6c8ff707411552e53a9b290f42b9fa881b38234",
       "bytes": 4949454016,
       "tensors": 95
     },
     {
       "file": "model-00004-of-00008.safetensors",
-      "sha256": "858fab295f100b897634b72824d01e01f368dc1d500de03916854d7b2769986c",
+      "sha256": "b90c3460f3cb1580b6ca663d70ec840a8fac2ae6ed043d9ee4450107fb744c56",
       "bytes": 4999819744,
       "tensors": 106
     },
     {
       "file": "model-00005-of-00008.safetensors",
-      "sha256": "0d31608093622df48c2284fa310170f9a1bfd127d7108bdacbc8106fcf296691",
+      "sha256": "ce7b416181248e1516e40805f9cfdb6e001c064308250fb92094563ce3d6c507",
       "bytes": 4999819776,
       "tensors": 106
     },
     {
       "file": "model-00006-of-00008.safetensors",
-      "sha256": "4356b837d798b2d3250cd067c15b7f5d56fd19a8b8605bb107fb9d727c1726d0",
+      "sha256": "34ddb4a6a1f67153e9ed34e23a5542a23c9032f0380c77c2c3b31caf8af5db89",
       "bytes": 4999819272,
       "tensors": 102
     },
     {
       "file": "model-00007-of-00008.safetensors",
-      "sha256": "003f5a6c4c727e5f560e16ad91ee2264ed138c7190108dfc9b68d1b8c8445545",
+      "sha256": "01bd1ff8c8a8a687033b6f6c0ac51c959235ec748547b0364a98508e2e444f12",
       "bytes": 4456023446,
       "tensors": 956
     },
     {
       "file": "model-00008-of-00008.safetensors",
-      "sha256": "299783fae2f540dfd723abbb09f8661b57122210d44b23a4c65f24b254542550",
+      "sha256": "e163845d238de6fc71a9990a2aaf0dea8514de1891a244d6fae9c96016bca92d",
       "bytes": 2155913464,
       "tensors": 2
     }