Upload TridaForDLM

README.md

@@ -0,0 +1,199 @@
+---
+library_name: transformers
+tags: []
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]

config.json

@@ -0,0 +1,69 @@
+{
+  "architectures": [
+    "TridaForDLM"
+  ],
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_trida.TridaConfig",
+    "AutoModel": "modeling_trida.TridaForDLM",
+    "AutoModelForCausalLM": "modeling_trida.TridaForDLM"
+  },
+  "bd_size": 4,
+  "bos_token_id": 0,
+  "dtype": "bfloat16",
+  "eos_token_id": 128001,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "initializer_range": 0.02,
+  "intermediate_size": 11008,
+  "layer_types": [
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention"
+  ],
+  "max_position_embeddings": 4096,
+  "max_window_layers": 28,
+  "model_type": "Trida",
+  "num_attention_heads": 32,
+  "num_hidden_layers": 32,
+  "num_key_value_heads": 32,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": null,
+  "rope_theta": 100000.0,
+  "sliding_window": null,
+  "tie_word_embeddings": false,
+  "transformers_version": "4.57.3",
+  "use_cache": true,
+  "use_sliding_window": false,
+  "vocab_size": 128256
+}

configuration_trida.py

@@ -0,0 +1,107 @@
+"""Trida model configuration"""
+
+from transformers.configuration_utils import PretrainedConfig
+try:
+    from transformers.configuration_utils import layer_type_validation
+except Exception:
+    def layer_type_validation(layer_types):
+        # Fallback for older/newer transformers without this helper
+        return
+from transformers.modeling_rope_utils import rope_config_validation
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class TridaConfig(PretrainedConfig):
+
+    model_type = "Trida"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    # Default tensor parallel plan for base model `Trida`
+    base_model_tp_plan = {
+        "layers.*.self_attn.q_proj": "colwise",
+        "layers.*.self_attn.k_proj": "colwise",
+        "layers.*.self_attn.v_proj": "colwise",
+        "layers.*.self_attn.o_proj": "rowwise",
+        "layers.*.mlp.gate_proj": "colwise",
+        "layers.*.mlp.up_proj": "colwise",
+        "layers.*.mlp.down_proj": "rowwise",
+    }
+    base_model_pp_plan = {
+        "embed_tokens": (["input_ids"], ["inputs_embeds"]),
+        "layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
+        "norm": (["hidden_states"], ["hidden_states"]),
+    }
+
+    def __init__(
+        self,
+        vocab_size=151936,
+        hidden_size=4096,
+        intermediate_size=22016,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=32,
+        hidden_act="silu",
+        max_position_embeddings=32768,
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        use_sliding_window=False,
+        sliding_window=4096,
+        max_window_layers=28,
+        layer_types=None,
+        attention_dropout=0.0,
+        bd_size=32,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.use_sliding_window = use_sliding_window
+        self.sliding_window = sliding_window if self.use_sliding_window else None
+        self.max_window_layers = max_window_layers
+
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self.attention_dropout = attention_dropout
+        self.bd_size = bd_size
+        # Validate the correctness of rotary position embeddings parameters
+        # BC: if there is a 'type' field, move it to 'rope_type'.
+        if self.rope_scaling is not None and "type" in self.rope_scaling:
+            self.rope_scaling["rope_type"] = self.rope_scaling["type"]
+        rope_config_validation(self)
+
+        self.layer_types = layer_types
+        if self.layer_types is None:
+            self.layer_types = [
+                "sliding_attention"
+                if self.sliding_window is not None and i >= self.max_window_layers
+                else "full_attention"
+                for i in range(self.num_hidden_layers)
+            ]
+        layer_type_validation(self.layer_types)
+
+        # head_dim can be passed via kwargs, otherwise default to 128
+        self.head_dim = kwargs.pop('head_dim', 128)
+
+        super().__init__(
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

generation_config.json

@@ -0,0 +1,8 @@
+{
+  "bos_token_id": 0,
+  "do_sample": true,
+  "eos_token_id": 128001,
+  "temperature": 0.6,
+  "top_p": 0.9,
+  "transformers_version": "4.57.3"
+}

model-00001-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e5e9a4589f938a297ead260acedab905c6dc4dc84e190628c58bdc70a560192f
+size 4917979000

model-00002-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:624ad919451158c5bc7229c6346dca79631abfedcd112f9bf128447ca92125c8
+size 4947390880

model-00003-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ae69bf99459c60fb06138805a4ed8cd8e4ad58a76e2d61220a15634947d23858
+size 4137880232

model-00004-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5f0a7f453efe5cea64bc17ef717b4177030167f0d724da0d8bda1fd2f89d4a29
+size 1050673280

model.safetensors.index.json

@@ -0,0 +1,299 @@
+{
+  "metadata": {
+    "total_parameters": 7526944768,
+    "total_size": 15053889536
+  },
+  "weight_map": {
+    "lm_head.weight": "model-00004-of-00004.safetensors",
+    "model.embed_tokens.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.input_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.input_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.10.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.10.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.10.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.10.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.10.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.10.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.10.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.10.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.10.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.14.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.14.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.14.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.14.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.14.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.14.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.14.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.14.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.14.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.15.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.15.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.15.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.15.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.15.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.15.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.15.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.15.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.15.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.16.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.16.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.16.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.16.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.16.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.16.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.16.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.16.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.16.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.17.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.17.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.17.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.17.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.17.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.17.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.17.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.17.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.17.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.18.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.18.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.18.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.18.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.18.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.18.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.18.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.18.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.18.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.19.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.19.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.19.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.19.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.19.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.19.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.19.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.19.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.19.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.2.input_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.2.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.2.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.2.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.2.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.2.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.2.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.2.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.2.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.20.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.20.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.20.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.20.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.20.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.20.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.20.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.20.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.20.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.21.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.21.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.21.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.21.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.21.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.21.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.21.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.21.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.21.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.22.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.22.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.22.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.22.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.22.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.22.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.22.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.22.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.22.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.23.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.23.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.23.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.23.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.23.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.23.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.23.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.23.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.23.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.24.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.24.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.24.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.24.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.24.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.24.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.24.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.24.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.24.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.25.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.25.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.25.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.25.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.25.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.25.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.25.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.25.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.25.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.26.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.26.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.26.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.26.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.26.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.26.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.26.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.26.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.26.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.27.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.27.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.27.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.27.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.27.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.27.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.27.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.27.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.27.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.28.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.28.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.28.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.28.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.28.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.28.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.28.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.28.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.28.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.29.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.29.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.29.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.29.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.29.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.29.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.29.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.29.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.29.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.3.input_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.3.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.3.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.3.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.3.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.3.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.3.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.3.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.3.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.30.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.30.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.30.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.30.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.30.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.30.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.30.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.30.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.30.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.31.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.31.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.31.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.31.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.31.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.31.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.31.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.31.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.31.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.4.input_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.5.input_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.5.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.5.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.5.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.5.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.5.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.5.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.5.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.5.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.6.input_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.6.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.6.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.6.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.6.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.6.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.6.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.6.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.6.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.7.input_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.7.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.7.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.7.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.7.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.7.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.7.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.7.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.7.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.8.input_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.8.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.8.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.8.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.8.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.8.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.8.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.8.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.8.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.9.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.9.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.9.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.9.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.9.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.9.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.9.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.9.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.9.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
+    "model.norm.weight": "model-00003-of-00004.safetensors"
+  }
+}

modeling_trida.py

@@ -0,0 +1,1008 @@
+from typing import Callable, Optional, Union
+from dataclasses import dataclass
+
+import torch
+from torch import nn
+import torch.nn.functional as F
+from functools import partial
+
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache
+from transformers.generation import GenerationMixin
+from transformers.integrations import use_kernel_forward_from_hub
+from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+from transformers.modeling_layers import GradientCheckpointingLayer
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+)
+from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from transformers.processing_utils import Unpack
+from transformers.utils import auto_docstring, can_return_tuple, logging
+from .configuration_trida import TridaConfig
+from torch.nn.attention.flex_attention import flex_attention, create_block_mask
+from einops import rearrange, repeat
+
+logger = logging.get_logger(__name__)
+
+
+@dataclass
+class CausalLMOutputWithPastAndBlockCache(CausalLMOutputWithPast):
+    block_past_key_values: Optional[Cache] = None
+    computed_attention_mask: Optional[torch.Tensor] = None
+
+@dataclass
+class BaseModelOutputWithPastAndBlockCache(BaseModelOutputWithPast):
+    block_past_key_values: Optional[Cache] = None
+    computed_attention_mask: Optional[torch.Tensor] = None
+
+
+@torch.compile(fullgraph=True, mode="max-autotune-no-cudagraphs")
+def fused_flex_attention(q, k, v, mask=None):
+    return flex_attention(q, k, v, block_mask=mask, enable_gqa=True)
+
+def block_diff_mask(b, h, q_idx, kv_idx, block_size=None, n=None):
+    """
+    Constructs the specialized block diffusion attention mask for training
+    composed of three masks:
+    - **Block Diagonal Mask (M_BD)**: Self-attention within noised blocks
+    - **Offset Block Causal Mask (M_OBC)**: Cross-attention for conditional context
+    - **Block Causal Mask (M_BC)**: Attention to update x0
+
+    Args:
+        b, h: Batch and head indices (ignored for mask logic).
+        q_idx, kv_idx: Query and Key indices.
+        seq_len: Total sequence length.
+        block_size: Defines the block structure.
+
+    Returns:
+        A boolean attention mask.
+    """
+    # Indicate whether token belongs to xt or x0
+    x0_flag_q = (q_idx >= n)
+    x0_flag_kv = (kv_idx >= n)
+
+    # Compute block indices
+    block_q = torch.where(x0_flag_q == 1,
+                        (q_idx - n) // block_size,
+                        q_idx // block_size)
+    block_kv = torch.where(x0_flag_kv == 1,
+                        (kv_idx - n) // block_size,
+                        kv_idx // block_size)
+
+    # **1. Block Diagonal Mask (M_BD) **
+    block_diagonal = (block_q == block_kv) & (x0_flag_q == x0_flag_kv)
+
+    # **2. Offset Block-Causal Mask (M_OBC) **
+    offset_block_causal = (
+    (block_q > block_kv)
+    & (x0_flag_kv == 1)
+    & (x0_flag_q == 0)
+    )
+
+    # **3. Block-Causal Mask (M_BC) **
+    block_causal = (block_q >= block_kv) & (x0_flag_kv == 1) & (x0_flag_q == 1)
+
+    # **4. Combine Masks **
+    return block_diagonal | offset_block_causal | block_causal
+
+def eval_block_diff_mask(q_idx, kv_idx, block_size=None):
+    # Compute block indices
+    block_q = q_idx // block_size
+    block_kv = kv_idx // block_size
+
+    return block_q >= block_kv
+
+class TridaMLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x):
+        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+        return down_proj
+
+
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`, *optional*):
+            Deprecated and unused.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+class TridaAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: TridaConfig, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
+        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
+        self.scaling = self.head_dim**-0.5
+        self.attention_dropout = config.attention_dropout
+        self.is_causal = True
+        self.q_proj = nn.Linear(config.hidden_size, config.num_attention_heads * self.head_dim, bias=False)
+        self.k_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=False)
+        self.v_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=False)
+        self.o_proj = nn.Linear(config.num_attention_heads * self.head_dim, config.hidden_size, bias=False)
+        self.sliding_window = config.sliding_window if config.layer_types[layer_idx] == "sliding_attention" else None
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor],
+        past_key_value: Optional[Cache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        update_past_key_values: Optional[bool] = False,
+        block_past_key_values: Optional[Cache] = None,
+        replace_position: Optional[int] = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]:
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, self.head_dim)
+
+        query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+        key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+
+        cos, sin = position_embeddings
+        # query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+        if self.training:
+            #split q into two parts
+            q_1 = query_states[:,:,:query_states.shape[2]//2]
+            q_2 = query_states[:,:,query_states.shape[2]//2:]
+            #split k into two parts
+            k_1 = key_states[:,:,:key_states.shape[2]//2]
+            k_2 = key_states[:,:,key_states.shape[2]//2:]
+            q_1, k_1 = apply_rotary_pos_emb(q_1, k_1, cos, sin)
+            q_2, k_2 = apply_rotary_pos_emb(q_2, k_2, cos, sin)
+            query_states = torch.cat((q_1, q_2), dim=-2)
+            key_states = torch.cat((k_1, k_2), dim=-2)
+        else:
+            query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        if block_past_key_values is not None:
+            if len(block_past_key_values) <= self.layer_idx:
+                cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+                key_states, value_states = block_past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)
+            else:
+                block_cache_key_states = block_past_key_values[self.layer_idx][0]
+                block_cache_value_states = block_past_key_values[self.layer_idx][1]
+                
+                block_cache_key_states[:, :, replace_position:replace_position+key_states.shape[2]] = key_states
+                block_cache_value_states[:, :, replace_position:replace_position+value_states.shape[2]] = value_states
+                key_states = block_cache_key_states
+                value_states = block_cache_value_states
+
+        if past_key_value is not None:
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            if update_past_key_values:
+                cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+                key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+            elif len(past_key_value) > self.layer_idx:
+                key_states = torch.cat((past_key_value[self.layer_idx][0], key_states), dim=-2)
+                value_states = torch.cat((past_key_value[self.layer_idx][1], value_states), dim=-2)
+
+        if self.training:
+            attn_output = fused_flex_attention(query_states, key_states, value_states, mask=attention_mask)
+            attn_output = attn_output.transpose(1, 2).contiguous()
+        else:
+            attention_interface = ALL_ATTENTION_FUNCTIONS["sdpa"]
+
+            attn_output, attn_weights = attention_interface(
+                self,
+                query_states,
+                key_states,
+                value_states,
+                attention_mask,
+                is_causal=False,
+                dropout=0.0 if not self.training else self.attention_dropout,
+                scaling=self.scaling,
+                sliding_window=self.sliding_window,  # main diff with Llama
+                **kwargs,
+            )
+
+        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+        return attn_output
+
+@use_kernel_forward_from_hub("RMSNorm")
+class TridaRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        RMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+    def extra_repr(self):
+        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
+
+
+class TridaDecoderLayer(GradientCheckpointingLayer):
+    def __init__(self, config: TridaConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+
+        self.self_attn = TridaAttention(config=config, layer_idx=layer_idx)
+
+        self.mlp = TridaMLP(config)
+        self.input_layernorm = TridaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = TridaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.attention_type = config.layer_types[layer_idx]
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
+        update_past_key_values: Optional[bool] = False,
+        use_block_cache: Optional[bool] = False,
+        block_past_key_values: Optional[Cache] = None,
+        replace_position: Optional[int] = None,
+        **kwargs
+    ) -> tuple[torch.Tensor]:
+        residual = hidden_states
+        hidden_states = self.input_layernorm(hidden_states)
+        # Self Attention
+        hidden_states = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            position_embeddings=position_embeddings,
+            update_past_key_values=update_past_key_values,
+            use_block_cache=use_block_cache,
+            block_past_key_values=block_past_key_values,
+            replace_position=replace_position,
+            **kwargs,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+        return hidden_states
+
+
+
+class TridaPreTrainedModel(PreTrainedModel):
+    config_class = TridaConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["TridaDecoderLayer"]
+    _skip_keys_device_placement = ["past_key_values"]
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    _supports_flex_attn = True
+    _supports_cache_class = True
+    _supports_quantized_cache = True
+    _supports_static_cache = True
+    _supports_attention_backend = True
+    _can_record_outputs = {
+        "hidden_states": TridaDecoderLayer,
+        "attentions": TridaAttention,
+    }
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, TridaRMSNorm):
+            module.weight.data.fill_(1.0)
+
+
+class TridaRotaryEmbedding(nn.Module):
+    def __init__(self, config: TridaConfig, device=None):
+        super().__init__()
+        # BC: "rope_type" was originally "type"
+        if hasattr(config, "rope_scaling") and isinstance(config.rope_scaling, dict):
+            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
+        else:
+            self.rope_type = "default"
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+
+        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = self.inv_freq
+
+    @torch.no_grad()
+    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
+    def forward(self, x, position_ids):
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
+        position_ids_expanded = position_ids[:, None, :].float()
+
+        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos() * self.attention_scaling
+            sin = emb.sin() * self.attention_scaling
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+
+class TridaModel(TridaPreTrainedModel):
+    def __init__(self, config: TridaConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+        self.bd_size = config.bd_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList(
+            [TridaDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.norm = TridaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = TridaRotaryEmbedding(config=config)
+        self.gradient_checkpointing = False
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+
+    def eval_mask(self, seqlen, block_size, cache_seq_len, input_ids=None, mask_id=128012,
+                  prev_eval_mask=None):
+        """
+        Creates attention mask for inference with:
+        1. Block-wise causal attention: tokens in current block attend to previous clean blocks
+        2. Within-block any-order causal attention: once a token is unmasked, its attention mask is fixed
+
+        Args:
+            seqlen: Current sequence length
+            block_size: Size of attention blocks
+            cache_seq_len: Length of cached sequence from previous blocks
+            input_ids: Current input token IDs [batch, seqlen]
+            mask_id: Token ID used for masked positions
+            prev_eval_mask: Previous attention mask to preserve frozen attention patterns
+        """
+        device = input_ids.device if input_ids is not None else 'cuda'
+        total_len = seqlen + cache_seq_len
+
+        q_indices = torch.arange(seqlen, device=device) + cache_seq_len
+        k_indices = torch.arange(total_len, device=device)
+
+        # Block-level causal mask: [seqlen, total_len]
+        block_q = q_indices[:, None] // block_size
+        block_kv = k_indices[None, :] // block_size
+        block_causal_mask = (block_q >= block_kv).unsqueeze(0)  # [1, seqlen, total_len]
+
+        if input_ids is None or mask_id is None:
+            return block_causal_mask
+
+        batch_size = input_ids.shape[0]
+
+        # Current mask state
+        is_mask_q = (input_ids == mask_id)  # [batch, seqlen]
+        is_mask_kv_cached = torch.zeros(batch_size, cache_seq_len, dtype=torch.bool, device=device)
+        is_mask_kv = torch.cat([is_mask_kv_cached, is_mask_q], dim=1)  # [batch, total_len]
+
+        # Diagonal (self-attention always allowed)
+        diagonal_mask = (q_indices[:, None] == k_indices[None, :]).unsqueeze(0)  # [1, seqlen, total_len]
+
+        if prev_eval_mask is not None:
+            # Any-order causal attention:
+            # - Currently unmasked tokens: use their frozen attention from prev_eval_mask
+            # - Still masked tokens: attend to all currently unmasked tokens + themselves
+
+            is_unmasked_kv = ~is_mask_kv  # [batch, total_len]
+
+            # For still masked queries: attend to all currently unmasked tokens + themselves
+            still_masked_q_mask = (block_causal_mask & is_unmasked_kv.unsqueeze(1)) | diagonal_mask
+
+            # Get previous mask (squeeze the head dimension)
+            prev_eval_mask_squeezed = prev_eval_mask.squeeze(1)  # [batch, seqlen, total_len]
+
+            # Expand is_mask_q for indexing: [batch, seqlen, total_len]
+            is_mask_q_exp = is_mask_q.unsqueeze(-1).expand(-1, -1, total_len)
+
+            # Build final mask:
+            # - Currently unmasked tokens: use their frozen attention from prev_eval_mask
+            # - Still masked tokens: use still_masked_q_mask
+            final_mask = torch.where(
+                is_mask_q_exp,
+                still_masked_q_mask,
+                prev_eval_mask_squeezed
+            )
+
+            return final_mask.unsqueeze(1)  # [batch, 1, seqlen, total_len]
+        else:
+            # First step or no previous mask info
+            # Masked tokens: attend to all unmasked tokens + themselves (to predict what to fill in)
+            # Unmasked tokens: block causal to other unmasked tokens + self
+
+            is_mask_q_exp = is_mask_q.unsqueeze(-1)  # [batch, seqlen, 1]
+            is_mask_kv_exp = is_mask_kv.unsqueeze(1)  # [batch, 1, total_len]
+
+            # Both masked and unmasked queries can see: unmasked keys within block causal + diagonal
+            final_mask = (block_causal_mask & ~is_mask_kv_exp) | diagonal_mask
+
+            return final_mask.unsqueeze(1)  # [batch, 1, seqlen, total_len]
+
+    def gen_mask(self, seqlen, block_size, B, H):
+        mask = create_block_mask(
+            partial(block_diff_mask, block_size=block_size, n=seqlen),
+            B=B, H=H, Q_LEN=seqlen*2, KV_LEN=seqlen*2)
+
+        return mask
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        update_past_key_values: Optional[bool] = False,
+        block_size: Optional[int] = 32,
+        use_block_cache: Optional[bool] = False,
+        block_past_key_values: Optional[Cache] = None,
+        replace_position: Optional[int] = None,
+        prev_attention_mask: Optional[torch.Tensor] = None,
+        mask_id: Optional[int] = 128012,
+        **kwargs
+    ) -> BaseModelOutputWithPast:
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache()
+
+        if use_block_cache and block_past_key_values is None:
+            block_past_key_values = DynamicCache()
+
+        if cache_position is None:
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            if self.training:
+                cache_position = torch.arange(
+                    past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1]//2, device=inputs_embeds.device
+                )
+            else:
+                if use_block_cache:
+                    block_start_position = past_seen_tokens+replace_position if replace_position is not None else past_seen_tokens
+                    cache_position = torch.arange(
+                        block_start_position, block_start_position + inputs_embeds.shape[1], device=inputs_embeds.device
+                    )
+                else:
+                    cache_position = torch.arange(
+                        past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1] if not self.training else inputs_embeds.shape[1]//2, device=inputs_embeds.device
+                    )
+
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+        
+        if self.training:
+            attention_mask = self.gen_mask(labels.shape[1], self.bd_size, labels.shape[0], self.config.num_attention_heads).to(device=inputs_embeds.device)
+        else:
+            # Always compute attention mask for proper any-order causal attention
+            cache_seq_len = past_key_values.get_seq_length() if past_key_values is not None else 0
+            attention_mask = self.eval_mask(
+                input_ids.shape[1],
+                block_size,
+                cache_seq_len,
+                input_ids=input_ids,
+                mask_id=mask_id,
+                prev_eval_mask=prev_attention_mask
+            ).to(device=inputs_embeds.device)
+
+        hidden_states = inputs_embeds
+
+        # create position embeddings to be shared across the decoder layers
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+
+        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
+            hidden_states = decoder_layer(
+                hidden_states,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_value=past_key_values,
+                use_cache=use_cache,
+                cache_position=cache_position,
+                position_embeddings=position_embeddings,
+                update_past_key_values=update_past_key_values,
+                use_block_cache=use_block_cache,
+                block_past_key_values=block_past_key_values,
+                replace_position=replace_position,
+                **kwargs,
+            )
+
+        hidden_states = self.norm(hidden_states)
+        return BaseModelOutputWithPastAndBlockCache(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values if use_cache else None,
+            block_past_key_values=block_past_key_values if use_block_cache else None,
+            computed_attention_mask=attention_mask if not self.training else None,
+        )
+
+
+class TridaForDLM(TridaPreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+    _tp_plan = {"lm_head": "colwise_rep"}
+    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = TridaModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    def _create_attention_mask(self, input_ids, seq_len, pad_token_id):
+        """
+        Create attention mask that marks padding positions as 0 (masked)
+        and real token positions as 1 (unmasked).
+
+        Args:
+            input_ids: Tensor of shape (batch_size, seq_length)
+            seq_len: Tensor of shape (batch_size,) with actual sequence lengths for each sample
+            pad_token_id: Token ID used for padding
+
+        Returns:
+            attention_mask: Tensor of shape (batch_size, seq_length) with 1 for real tokens, 0 for padding
+        """
+        batch_size, seq_length = input_ids.shape
+        attention_mask = torch.ones((batch_size, seq_length), dtype=torch.long, device=input_ids.device)
+
+        # Mark padding positions as 0
+        for i in range(batch_size):
+            # Padding can occur after the actual sequence length
+            if seq_len[i] < seq_length:
+                attention_mask[i, seq_len[i]:] = 0
+
+        return attention_mask
+
+    @can_return_tuple
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
+        update_past_key_values: Optional[bool] = False,
+        block_size: Optional[int] = 32,
+        use_block_cache: Optional[bool] = False,
+        block_past_key_values: Optional[Cache] = None,
+        replace_position: Optional[int] = None,
+        mask_id: Optional[int] = 128012,
+        prev_attention_mask: Optional[torch.Tensor] = None,
+        **kwargs
+    ) -> CausalLMOutputWithPastAndBlockCache:
+
+        if self.training:
+            original_labels = labels.clone()
+            original_input_ids = input_ids.clone()
+
+            noisy_input_ids = input_ids.clone()
+
+            input_ids = input_ids.reshape(input_ids.shape[0] * input_ids.shape[1] // self.model.bd_size, self.model.bd_size)
+            b, l = input_ids.shape
+            t = torch.rand((b,), device=input_ids.device)
+            eps=1e-3
+            p_mask = (1 - eps) * t + eps
+            p_mask = p_mask[:, None].repeat(1, l)
+
+            mask_indices = torch.rand((b, l), device=input_ids.device) < p_mask
+            x_t = torch.where(mask_indices, mask_id, input_ids).reshape(labels.shape)
+            noisy_input_ids[labels != -100] = x_t[labels != -100]
+            mask = (noisy_input_ids != mask_id)
+            labels[mask] = -100
+            input_ids = torch.cat([noisy_input_ids, input_ids.reshape(labels.shape)], dim=1)
+
+            complementary_noisy_input_ids = original_input_ids.clone()
+            complementary_labels = original_labels.clone()
+
+            complementary_input_ids = original_input_ids.reshape(original_input_ids.shape[0] * original_input_ids.shape[1] // self.model.bd_size, self.model.bd_size)
+
+            complementary_mask_indices = ~mask_indices
+            complementary_x_t = torch.where(complementary_mask_indices, mask_id, complementary_input_ids).reshape(labels.shape)
+            complementary_noisy_input_ids[complementary_labels != -100] = complementary_x_t[complementary_labels != -100]
+            complementary_mask = (complementary_noisy_input_ids != mask_id)
+            complementary_labels[complementary_mask] = -100
+            complementary_input_ids = torch.cat([complementary_noisy_input_ids, complementary_input_ids.reshape(complementary_labels.shape)], dim=1)
+
+            input_ids = torch.cat([input_ids, complementary_input_ids], dim=0)
+            labels = torch.cat([labels, complementary_labels], dim=0)
+
+        outputs: BaseModelOutputWithPastAndBlockCache = self.model(
+            input_ids=input_ids,
+            labels=labels,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            update_past_key_values=update_past_key_values,
+            block_size=block_size,
+            use_block_cache=use_block_cache,
+            block_past_key_values=block_past_key_values,
+            replace_position=replace_position,
+            prev_attention_mask=prev_attention_mask,
+            mask_id=mask_id,
+            **kwargs,
+        )
+
+        hidden_states = outputs.last_hidden_state
+        if self.training:
+            hidden_states = hidden_states[:, :hidden_states.shape[1]//2, :]
+        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+
+        loss = None
+        if labels is not None:
+            # Standard causal LM loss with automatic label shifting (logits[i] predicts labels[i+1])
+            # This is correct even for block diffusion training
+            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)
+
+        return CausalLMOutputWithPastAndBlockCache(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            block_past_key_values=outputs.block_past_key_values,
+            computed_attention_mask=outputs.computed_attention_mask,
+        )
+
+    def _has_stop_token(self, tokens, stop_token):
+        """Check if any batch has the stop token. Returns (any_has_stop, per_batch_mask)."""
+        has_stop = (tokens == stop_token).any(dim=-1)  # [batch]
+        return has_stop.any().item(), has_stop
+
+    def _get_first_stop_token_idx(self, tokens, stop_token):
+        """Get the index of the first stop token for each batch. Returns -1 if not found."""
+        batch_size, seq_len = tokens.shape
+        # Find positions where stop_token occurs
+        is_stop = (tokens == stop_token)  # [batch, seq]
+        # For each batch, find the first occurrence (or seq_len if not found)
+        # Use argmax on the boolean tensor - it returns first True index, or 0 if all False
+        has_stop = is_stop.any(dim=-1)  # [batch]
+        first_idx = is_stop.float().argmax(dim=-1)  # [batch]
+        # Set to -1 for batches without stop token
+        first_idx = torch.where(has_stop, first_idx, torch.tensor(-1, device=tokens.device))
+        return first_idx  # [batch]
+
+    def _get_per_sequence_stop_status(self, tokens, prompt_length, stop_token, mask_id):
+        """
+        Check stop condition per sequence.
+        Returns a boolean tensor [batch] where True means the sequence should stop.
+        A sequence stops when it has a stop token with no mask tokens before it.
+        """
+        batch_size = tokens.shape[0]
+        generated = tokens[:, prompt_length:]
+
+        # Get stop token positions
+        first_stop_idx = self._get_first_stop_token_idx(generated, stop_token)  # [batch]
+        has_stop = first_stop_idx >= 0  # [batch]
+
+        # For each sequence, check if there are mask tokens before the stop token
+        should_stop = torch.zeros(batch_size, dtype=torch.bool, device=tokens.device)
+        for b in range(batch_size):
+            if has_stop[b]:
+                # Check if any mask tokens exist before the stop token
+                if (generated[b, :first_stop_idx[b]] == mask_id).sum() == 0:
+                    should_stop[b] = True
+
+        return should_stop
+
+    @torch.no_grad()
+    def generate(
+        self,
+        input_ids,
+        max_new_tokens,
+        mask_id=128012,
+        threshold=1,
+        small_block_size=8,
+        block_size=32,
+        stop_token=128001,
+        stopping_criteria=None,
+        top_p=0.95,
+        temperature=0,
+        use_block_cache=False,
+        pad_token_id=128004,
+        return_dict_in_generate=False,
+        **kwargs
+    ):
+        batch_size = input_ids.shape[0]
+        num_blocks = max_new_tokens // block_size
+        original_input_length = input_ids.shape[1]
+
+        # Track which sequences are finished (have generated stop token with no masks before it)
+        finished = torch.zeros(batch_size, dtype=torch.bool, device=self.device)
+
+        if input_ids.shape[1] > block_size:
+            output = self.forward(input_ids=input_ids[:, :(input_ids.shape[1] // block_size * block_size)], use_cache=True, update_past_key_values=True, block_size=block_size)
+            logits, past_key_values = output.logits, output.past_key_values
+            if input_ids.shape[1] % block_size == 0:
+                next_token = logits[:, -1:, :].argmax(dim=-1)
+                input_ids = torch.cat([input_ids, next_token], dim=1)
+        else:
+            past_key_values = None
+
+        num_small_blocks = block_size // small_block_size
+
+        for block_idx in range(num_blocks):
+            # Check if all sequences are finished
+            if finished.all():
+                break
+
+            prompt_length = input_ids.shape[1]
+            # Initialize x_init with mask_id
+            x_init = mask_id * torch.ones((batch_size, block_size - prompt_length % block_size), device=self.device, dtype=torch.long)
+            x_init = torch.cat([input_ids, x_init], dim=1)
+
+            x_t = x_init.clone()
+            block_past_key_values = None
+            block_attention_mask = None  # Track attention mask within block for any-order causal
+
+            while True:
+                # Update finished status per sequence
+                newly_finished = self._get_per_sequence_stop_status(x_t, prompt_length, stop_token, mask_id)
+                finished = finished | newly_finished
+
+                # Check if all sequences are finished
+                if finished.all():
+                    break
+
+                # Only consider mask positions for unfinished sequences
+                mask_idx = (x_t[:, -block_size:] == mask_id)
+                # Zero out mask_idx for finished sequences (they don't need more unmasking)
+                mask_idx = mask_idx & (~finished.unsqueeze(-1))
+
+                # Decode a complete block, update cache, and generate the next token
+                if mask_idx.sum() == 0:
+                    output = self.forward(input_ids=x_t[:, -block_size:], use_cache=True, past_key_values=past_key_values, update_past_key_values=True, block_size=block_size)
+                    logits, past_key_values = output.logits, output.past_key_values
+                    next_token = logits[:, -1:, :].argmax(dim=-1)
+                    # Only append next token for unfinished sequences; use pad for finished
+                    next_token = torch.where(finished.unsqueeze(-1), pad_token_id, next_token)
+                    x_t = torch.cat([x_t, next_token], dim=1)
+                    break
+
+                for small_block_idx in range(num_small_blocks):
+                    small_block_start_idx = small_block_idx * small_block_size
+                    small_block_end_idx = small_block_start_idx + small_block_size
+
+                    start = -block_size + small_block_start_idx
+                    end = None if block_size == small_block_end_idx else -block_size + small_block_end_idx
+
+                    while True:
+                        # Recompute mask_idx considering finished sequences
+                        mask_idx = (x_t[:, -block_size:] == mask_id)
+                        mask_idx = mask_idx & (~finished.unsqueeze(-1))
+
+                        if mask_idx[:, start:end].sum() == 0:
+                            break
+
+                        # Update finished status
+                        newly_finished = self._get_per_sequence_stop_status(x_t, prompt_length, stop_token, mask_id)
+                        finished = finished | newly_finished
+
+                        if finished.all():
+                            break
+
+                        if use_block_cache:
+                            if block_past_key_values is None or (x_t[:, -block_size+small_block_start_idx] == mask_id).any():
+                                output = self.forward(input_ids=x_t[:, -block_size:], use_cache=True, past_key_values=past_key_values, update_past_key_values=False, use_block_cache=True, prev_attention_mask=block_attention_mask, mask_id=mask_id)
+                                logits, block_past_key_values = output.logits, output.block_past_key_values
+                                block_attention_mask = output.computed_attention_mask  # Track attention mask for any-order causal
+                                logits = torch.cat([logits[:, :1, :], logits[:, :-1, :]], dim=1)
+                                logits = logits[:, start:end]
+                            else:
+                                output = self.forward(input_ids=x_t[:,start:end], use_cache=True, past_key_values=past_key_values, update_past_key_values=False, use_block_cache=True, block_past_key_values=block_past_key_values, replace_position=small_block_start_idx, prev_attention_mask=block_attention_mask, mask_id=mask_id)
+                                logits = output.logits
+                                block_attention_mask = output.computed_attention_mask  # Track attention mask for any-order causal
+                                logits = torch.cat([logits[:, :1, :], logits[:, :-1, :]], dim=1)
+                        else:
+                            output = self.forward(input_ids=x_t[:, -block_size:], use_cache=True, past_key_values=past_key_values, update_past_key_values=False, prev_attention_mask=block_attention_mask, mask_id=mask_id)
+                            logits = output.logits
+                            block_attention_mask = output.computed_attention_mask  # Track attention mask for any-order causal
+                            logits = torch.cat([logits[:, :1, :], logits[:, :-1, :]], dim=1)
+                            logits = logits[:, start:end]
+
+                        # Allow stop_token to be generated naturally (don't block it)
+                        x_1, p_1t = self.sample_with_top_p(logits, top_p=top_p, temperature=temperature, stop_token_id=None)
+                        # Select tokens with probability greater than threshold from p_1t
+                        x1_p = torch.squeeze(torch.gather(p_1t, dim=-1, index=torch.unsqueeze(x_1, -1)), -1)
+                        x1_p = torch.where(mask_idx[:, start:end], x1_p, -torch.inf)
+
+                        unmask_idx = (x1_p > threshold)
+                        max_prob_idx = x1_p.argmax(dim=-1)
+                        unmask_idx[torch.arange(x_1.shape[0], device=self.device), max_prob_idx] = True
+                        unmask_idx = unmask_idx & mask_idx[:, start:end]
+
+                        # Don't update tokens for finished sequences
+                        unmask_idx = unmask_idx & (~finished.unsqueeze(-1))
+
+                        x_t[:, start:end][unmask_idx] = x_1[unmask_idx]
+
+            input_ids = x_t
+
+        # Per-sequence truncation: truncate each sequence at its stop token
+        # Pad shorter sequences to maintain consistent tensor shape
+        first_stop_idx = self._get_first_stop_token_idx(input_ids[:, original_input_length:], stop_token)
+
+        # Calculate the max output length (considering sequences without stop token)
+        max_output_len = input_ids.shape[1]
+        has_stop = first_stop_idx >= 0
+        if has_stop.any():
+            # For sequences with stop token: original_input_length + stop_idx + 1 (include stop token)
+            # For sequences without: keep full length
+            output_lens = torch.where(
+                has_stop,
+                original_input_length + first_stop_idx + 1,
+                torch.tensor(input_ids.shape[1], device=self.device)
+            )
+            max_output_len = output_lens.max().item()
+
+        # Truncate to max_output_len
+        input_ids = input_ids[:, :max_output_len]
+
+        # Replace tokens after each sequence's stop token with pad_token_id
+        for b in range(batch_size):
+            if has_stop[b]:
+                stop_pos = original_input_length + first_stop_idx[b].item() + 1
+                if stop_pos < input_ids.shape[1]:
+                    input_ids[b, stop_pos:] = pad_token_id
+
+        return input_ids
+
+    def sample_with_top_p(self, logits, top_p=0.95, temperature=1.0, mask_token_id=128012, pad_token_id=128004, stop_token_id=None):
+        # Mask out special tokens to prevent them from being generated during unmasking
+        # mask_token and pad_token should never be generated as actual output
+        # stop_token_id can be optionally masked during intra-block unmasking to prevent premature stopping
+        logits = logits.clone()
+        logits[..., mask_token_id] = -float('inf')
+        logits[..., pad_token_id] = -float('inf')
+        if stop_token_id is not None:
+            logits[..., stop_token_id] = -float('inf')
+
+        # Calculate probabilities
+        if temperature > 0:
+            scaled_logits = logits / temperature
+        else:
+            p_1t = torch.softmax(logits, dim=-1)
+            x_1 = p_1t.argmax(dim=-1)
+            return x_1, p_1t
+
+        probs = F.softmax(scaled_logits, dim=-1)
+
+        sorted_probs, sorted_indices = torch.sort(probs, descending=True)
+        cumulative_probs = torch.cumsum(sorted_probs, dim=-1)
+
+        sorted_indices_to_remove = cumulative_probs > top_p
+        sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
+        sorted_indices_to_remove[..., 0] = 0
+
+        indices_to_remove = torch.zeros_like(probs, dtype=torch.bool).scatter_(
+            dim=-1, index=sorted_indices, src=sorted_indices_to_remove
+        )
+        
+        probs[indices_to_remove] = 0
+
+        # Renormalize so that the probabilities of remaining tokens sum to 1
+        # Add a small epsilon value to prevent division by zero
+        probs_sum = torch.sum(probs, dim=-1, keepdim=True)
+        probs_sum = torch.clamp(probs_sum, min=1e-10)
+        normalized_probs = probs / probs_sum
+
+        p_1t = normalized_probs
+        # Handle arbitrary batch sizes: reshape to [batch * seq, vocab], sample, reshape back
+        batch_size, seq_len, vocab_size = p_1t.shape
+        p_1t_flat = p_1t.view(-1, vocab_size)
+        x_1 = torch.multinomial(p_1t_flat, num_samples=1).view(batch_size, seq_len)
+
+        return x_1, p_1t