MilliyLM-5B: instruction-tuned Uzbek chat model (SFT of NeuronAI-5B-Base)

README.md

@@ -0,0 +1,324 @@
+---
+license: apache-2.0
+language:
+  - uz
+  - en
+  - ru
+pipeline_tag: text-generation
+tags:
+  - uzbek
+  - o'zbek
+  - chat
+  - instruction-tuned
+base_model: kmamaroziqov/NeuronAI-5B-Base
+library_name: transformers
+---
+
+# MilliyLM-5B
+
+An instruction-tuned Uzbek chat model, supervised fine-tuned from
+[`kmamaroziqov/NeuronAI-5B-Base`](https://huggingface.co/kmamaroziqov/NeuronAI-5B-Base).
+
+MilliyLM-5B is a **chat and text-classification model**. It follows Uzbek instructions
+reliably, writes fluent Uzbek in both Latin and Cyrillic script, and is strong on
+sentiment and news classification. It is **not** a knowledge model: on multiple-choice
+knowledge benchmarks it performs at chance. Read the
+[Evaluation](#evaluation) and [Limitations](#limitations) sections before using it —
+they are specific about what works and what does not.
+
+| | |
+|---|---|
+| Parameters | 5.17 B |
+| Architecture | `NeuronLMForCausalLM` (custom, ships with the repo) |
+| Layers / hidden | 36 / 3584 |
+| Attention | GQA, 28 query heads : 4 KV heads, head_dim 128, QK-norm |
+| Position encoding | RoPE, θ = 500000 |
+| Context length | 4096 tokens |
+| Vocabulary | 48,000 (BPE) |
+| Embeddings | untied |
+| Weights dtype | bfloat16 |
+| Languages | Uzbek (Latin + Cyrillic), English, Russian |
+
+---
+
+## Quick start
+
+The architecture is custom, so `trust_remote_code=True` is **required** — the modeling
+code ships inside this repository.
+
+```python
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+model_id = "NeuronUz/MilliyLM-5B"
+
+tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
+model = AutoModelForCausalLM.from_pretrained(
+    model_id,
+    trust_remote_code=True,
+    dtype=torch.bfloat16,       # weights are bf16; do not load in fp32
+    device_map="cuda",
+).eval()
+
+messages = [{"role": "user", "content": "O'zbekistonning poytaxti qaysi shahar?"}]
+inputs = tokenizer.apply_chat_template(
+    messages,
+    add_generation_prompt=True,
+    return_tensors="pt",
+    return_dict=True,
+).to(model.device)
+
+with torch.no_grad():
+    out = model.generate(
+        **inputs,
+        max_new_tokens=256,
+        do_sample=False,           # greedy; see Generation settings below
+        eos_token_id=5,            # <|im_end|>  -- NOT the config's </s>
+        pad_token_id=3,            # <pad>
+    )
+
+print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
+```
+
+```
+Oʻzbekistonning poytaxti - Toshkent.
+```
+
+### Chat template
+
+The model uses ChatML. `tokenizer.apply_chat_template` applies it for you; the raw form is:
+
+```
+<|im_start|>system
+{system}<|im_end|>
+<|im_start|>user
+{user}<|im_end|>
+<|im_start|>assistant
+{assistant}<|im_end|>
+```
+
+A system turn is optional. Uzbek-language system prompts work best — that is what the
+model was trained with.
+
+### Generation settings
+
+| setting | value | why |
+|---|---|---|
+| `eos_token_id` | **5** (`<|im_end|>`) | The turn terminator. Token id 1 (`</s>`) is the *pretraining* EOS and never appears in chat data — using it means generation runs to `max_new_tokens`. |
+| `pad_token_id` | 3 (`<pad>`) | |
+| `do_sample` | `False` for classification/extraction; `True`, `temperature≈0.7`, `top_p≈0.9` for open chat | Every benchmark below was measured greedy. |
+| `dtype` | `torch.bfloat16` | Trained in bf16. |
+
+Memory: ~10.5 GB for weights in bf16, so a single 16 GB GPU is enough for inference.
+
+### Serving
+
+**vLLM and SGLang cannot load this model.** They reimplement each architecture
+internally rather than executing a repository's Python, and `NeuronLMForCausalLM` is not
+in their model registries — `trust_remote_code` only covers the config and tokenizer
+there. Use the `transformers` backend, or convert the weights (the architecture is
+Qwen3-equivalent apart from *fused* `qkv_proj` / `gate_up_proj` and `out_proj` naming;
+splitting those tensors and renaming to the Qwen3 layout yields a checkpoint vLLM will
+serve).
+
+---
+
+## Training
+
+Supervised fine-tuning only — no continued pretraining was performed on top of the base.
+
+| | |
+|---|---|
+| Method | Full-parameter SFT (no LoRA) |
+| Data | 190,939 instruction/chat examples |
+| Epochs | 3 (17,701 steps); released checkpoint is from epoch 2.75 |
+| Effective batch | 32 (micro-batch 8 × grad-accum 4) |
+| Sequence length | 2048 |
+| Optimizer | AdamW fused, β₁ 0.9, β₂ 0.95, weight decay 0.0 |
+| LR schedule | 1e-5, cosine, 3% warmup, grad-norm clip 1.0 |
+| Precision | bf16 mixed precision, gradient checkpointing |
+| Hardware | 1 × NVIDIA RTX PRO 6000 Blackwell (96 GB), ~10 h |
+
+### Data composition
+
+| slice | rows | share |
+|---|---:|---:|
+| Uzbek instruction/chat backbone (curated + filtered) | 169,919 | 89.0% |
+| Uzbek **Cyrillic** chat (transliterated) | 12,000 | 6.3% |
+| Russian-instructed translation | 5,000 | 2.6% |
+| Latin ↔ Cyrillic script conversion | 3,000 | 1.6% |
+| Cyrillic identity/social | 1,020 | 0.5% |
+
+The backbone mixes general Uzbek assistant data, benchmark-format task data
+(MCQ, classification, spelling), English↔Uzbek translation pairs, and an English
+retention slice. Third-person rubric-grading text was filtered out of the backbone
+before training.
+
+The Cyrillic and Russian slices exist because the base model reads and writes Uzbek
+Cyrillic *better* than Latin (bits-per-byte 0.2288 vs 0.3868) yet had almost no Cyrillic
+chat behaviour attached to it, and because Russian-language instructions were nearly
+absent. Checkpoint selection was done by running the full benchmark suite on all 12
+saved checkpoints, not by held-out loss — held-out loss was flat (1.784–1.796) across
+the last two epochs while benchmark scores were still moving.
+
+---
+
+## Evaluation
+
+Full public benchmark suite, greedy decoding, `transformers` backend, seed 42, complete
+test sets (no subsampling). Scores are accuracy unless noted.
+
+### Uzbek benchmarks
+
+| benchmark | n | score | invalid rate |
+|---|---:|---:|---:|
+| uzlib (Uzbek linguistic MCQ) | 1,861 | 0.2875 | 0.0000 |
+| TUMLU-Uzbek (Uzbek MMLU) | 700 | 0.3286 | 0.0000 |
+| MMLU-Uz (translated MMLU) | 14,042 | 0.2584 | 0.0000 |
+| News topic classification (10-way) | 96,970 | **0.6531** | 0.0000 |
+| Sentiment (binary) | 10,000 | **0.9259** | 0.0001 |
+
+Random baselines: 0.25 for the 4-way MCQ tasks, 0.10 for news, 0.50 for sentiment.
+
+### English
+
+| benchmark | n | score | invalid rate |
+|---|---:|---:|---:|
+| MMLU (English) | 14,042 | 0.2619 | 0.0000 |
+
+### Translation (FLORES+)
+
+| direction | n | BLEU | COMET | length ratio |
+|---|---:|---:|---:|---:|
+| English → Uzbek | 2,009 | 5.17 | 0.7397 | 1.018 |
+| Uzbek → English | 2,009 | 1.83 | 0.5376 | 1.229 |
+
+### uzlib, per split
+
+| split | n | score |
+|---|---:|---:|
+| fill_in | 52 | 0.3077 |
+| correct_word (orthography) | 1,501 | 0.3011 |
+| meaning_in_context | 72 | 0.2639 |
+| meaning | 236 | 0.2034 |
+
+### News, per class
+
+| class | n | score |
+|---|---:|---:|
+| Sport | 16,113 | 0.8743 |
+| class 2 | 5,177 | 0.7309 |
+| class 4 | 2,405 | 0.7081 |
+| class 0 | 29,500 | 0.6794 |
+| class 1 | 10,755 | 0.6596 |
+| class 5 | 3,505 | 0.6579 |
+| class 7 | 1,987 | 0.6548 |
+| class 8 | 1,784 | 0.5667 |
+| class 9 | 11,732 | 0.5124 |
+| Oila va Jamiyat (Family & Society) | 14,012 | 0.4273 |
+
+### Comparison with prior SFTs of the same base
+
+Same benchmark suite, same conditions.
+
+| benchmark | **MilliyLM-5B** | NeuronAI-5B-v4 | NeuronAI-5B (v1) |
+|---|---:|---:|---:|
+| uzlib | **0.2875** | 0.2708 | 0.2638 |
+| TUMLU-Uz | **0.3286** | 0.2100 | 0.1914 |
+| MMLU-Uz | **0.2584** | 0.2136 | 0.2154 |
+| MMLU (English) | **0.2619** | 0.2142 | 0.2196 |
+| News | **0.6531** | 0.1834 | 0.1542 |
+| Sentiment | **0.9259** | 0.1186 | 0.2647 |
+| FLORES en→uz BLEU | 5.17 | **7.68** | 7.37 |
+| FLORES uz→en BLEU | 1.83 | **4.21** | 5.15 |
+
+Most of the classification gain comes from **format compliance** rather than raw
+capability: the earlier models emitted unparseable answers on 63–78% of sentiment items
+and 6–13% of TUMLU items, while MilliyLM-5B's invalid rate is ≤0.0001 across every task.
+Translation is the one axis where the earlier models are better — see Limitations.
+
+### Contamination check
+
+44.1% of the sentiment evaluation set also appears in the training data, because the
+benchmark scores the dataset's `train` split and the task-format training rows were drawn
+from the same pool. This was tested rather than assumed:
+
+| slice | n | score |
+|---|---:|---:|
+| items seen in training | 1,200 | 0.9342 |
+| items not seen (exact match excluded) | 1,200 | 0.9300 |
+| items not seen (exact **and** normalized match excluded) | 1,500 | 0.9347 |
+
+Performance on strictly unseen data is identical to performance on memorized data, so
+the sentiment score reflects genuine capability. The news benchmark has **zero** overlap
+with training data.
+
+---
+
+## Limitations
+
+**Multiple-choice knowledge tasks perform at chance.** uzlib, MMLU-Uz and MMLU-English
+all sit within noise of their 0.25 random baseline, across roughly 30,000 questions.
+Invalid rates near zero mean the model answers in the correct format every time and is
+still wrong — this is missing knowledge, not broken parsing. The base model completed a
+single pretraining epoch, and supervised fine-tuning cannot add facts that were never
+learned. **Do not use this model for factual question answering, exams, or retrieval-free
+knowledge tasks.** TUMLU-Uzbek at 0.3286 is the only MCQ result above chance, and its
+700-item sample gives it a ±3.5% confidence interval.
+
+**Uzbek → English translation is weak and regressed against the base's earlier SFTs.**
+BLEU 1.83 with a 1.229 length ratio and 12.5% unigram precision means the model
+over-generates English that mostly does not match the reference. English → Uzbek is
+usable (COMET 0.7397) but not competitive with dedicated translation systems.
+
+**Script conversion does not work despite being trained for it.** Asked to transliterate
+Latin Uzbek to Cyrillic, the model frequently returns the input unchanged. The 3,000-row
+slice was too small.
+
+**The Cyrillic slice was machine-transliterated, and its artifacts are visible in
+output.** Loanwords and brand names inside Cyrillic text can come out mangled
+(e.g. `Facebook` → `Факебоок`), and occasional single Cyrillic characters leak into
+Latin words. Cyrillic *chat* is coherent and does not degenerate, but Cyrillic
+*orthography* is less reliable than Latin.
+
+**Self-identification.** The identity training data names the model "NeuronAI 5B", so
+asked who it is, it answers with that name rather than "MilliyLM-5B".
+
+**News classification is uneven.** The "Oila va Jamiyat" (Family & Society) class scores
+0.4273 across 14,012 items — a semantically diffuse catch-all the model handles poorly,
+against 0.8743 for the lexically distinctive Sport class.
+
+**Safety.** No safety alignment, RLHF, or red-teaming was performed. The model has no
+refusal training beyond what the instruction data incidentally contains. It can produce
+incorrect, biased, or unsafe content, and — given the benchmark results above — will
+state false facts fluently and confidently. Evaluate it for your own use case before
+deploying it anywhere user-facing.
+
+---
+
+## Intended use
+
+**Suitable for:** Uzbek-language chat and assistance; text classification (sentiment,
+topic); Uzbek text generation and rewriting in Latin or Cyrillic; English → Uzbek
+translation where approximate meaning suffices; a base for further fine-tuning.
+
+**Not suitable for:** factual question answering or anything knowledge-intensive;
+exam-style multiple choice; Uzbek → English translation; script transliteration; any
+application where a confidently-stated wrong fact causes harm (medical, legal, financial
+advice).
+
+## License
+
+Apache 2.0, inherited from the base model. Training data licensing follows the sources
+of the underlying public datasets.
+
+## Citation
+
+```bibtex
+@misc{milliylm5b,
+  title  = {MilliyLM-5B: an instruction-tuned Uzbek language model},
+  author = {NeuronUz},
+  year   = {2026},
+  url    = {https://huggingface.co/NeuronUz/MilliyLM-5B}
+}
+```

attention.py

@@ -0,0 +1,308 @@
+from __future__ import annotations
+
+from typing import Any
+
+import torch
+import torch.nn.functional as F
+from torch import Tensor, nn
+from transformers import Cache
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS
+
+from .configuration_neuron_lm import NeuronLMConfig
+from .layers import RMSNorm
+from .rotary import apply_rotary_pos_emb
+
+__all__ = ["NeuronLMAttention"]
+
+
+def _repeat_kv(hidden_states: Tensor, repeats: int) -> Tensor:
+    if repeats == 1:
+        return hidden_states
+    return hidden_states.repeat_interleave(repeats, dim=1)
+
+
+def eager_attention_forward(
+    module: NeuronLMAttention,
+    query: Tensor,
+    key: Tensor,
+    value: Tensor,
+    attention_mask: Tensor | None,
+    *,
+    scaling: float,
+    dropout: float = 0.0,
+    **_: Any,
+) -> tuple[Tensor, Tensor]:
+    """Numerically clear GQA reference used for attention-weight outputs."""
+
+    key = _repeat_kv(key, module.num_key_value_groups)
+    value = _repeat_kv(value, module.num_key_value_groups)
+    attention_weights = (
+        torch.matmul(
+            query,
+            key.transpose(-2, -1),
+        )
+        * scaling
+    )
+
+    fully_masked: Tensor | None = None
+    if attention_mask is not None:
+        if attention_mask.dtype == torch.bool:
+            fully_masked = ~attention_mask.any(
+                dim=-1,
+                keepdim=True,
+            )
+            attention_weights = attention_weights.masked_fill(
+                ~attention_mask,
+                torch.finfo(attention_weights.dtype).min,
+            )
+        else:
+            minimum = torch.finfo(attention_mask.dtype).min
+            fully_masked = (
+                torch.isneginf(attention_mask) | (attention_mask == minimum)
+            ).all(dim=-1, keepdim=True)
+            attention_weights = attention_weights + attention_mask
+
+    if fully_masked is not None:
+        attention_weights = attention_weights.masked_fill(
+            fully_masked,
+            0.0,
+        )
+
+    attention_weights = F.softmax(
+        attention_weights,
+        dim=-1,
+        dtype=torch.float32,
+    ).to(query.dtype)
+    attention_weights = torch.nan_to_num(
+        attention_weights,
+        nan=0.0,
+    )
+    if fully_masked is not None:
+        attention_weights = attention_weights.masked_fill(
+            fully_masked,
+            0.0,
+        )
+    attention_weights = F.dropout(
+        attention_weights,
+        p=dropout,
+        training=module.training,
+    )
+    attention_output = torch.matmul(attention_weights, value)
+    return (
+        attention_output.transpose(1, 2).contiguous(),
+        attention_weights,
+    )
+
+
+class NeuronLMAttention(nn.Module):
+    """Fused, bias-free GQA using a checkpoint-stable ``[Q, K, V]`` layout."""
+
+    def __init__(
+        self,
+        config: NeuronLMConfig,
+        layer_idx: int = 0,
+    ) -> None:
+        super().__init__()
+
+        if type(layer_idx) is not int or layer_idx < 0:
+            raise ValueError(
+                f"layer_idx must be a non-negative integer, got {layer_idx!r}"
+            )
+
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.num_key_value_heads = config.num_key_value_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+        self.head_dim = config.head_dim
+        self.scaling = self.head_dim**-0.5
+        self.attention_dropout = config.attention_dropout
+        self.layer_idx = layer_idx
+        self.is_causal = True
+
+        self.query_size = self.num_heads * self.head_dim
+        self.key_value_size = self.num_key_value_heads * self.head_dim
+
+        # State-dict contract: rows are Q, then K, then V.
+        self.qkv_proj = nn.Linear(
+            in_features=self.hidden_size,
+            out_features=config.qkv_projection_size,
+            bias=False,
+        )
+        self.out_proj = nn.Linear(
+            in_features=self.query_size,
+            out_features=self.hidden_size,
+            bias=False,
+        )
+
+        # Per-head normalization of queries and keys before RoPE, as in
+        # Qwen3 / OLMo-2 / Gemma-3. Bounds the growth of q.k during long bf16
+        # runs, which depth-scaled initialization does not address: init
+        # controls the residual stream at step 0, while attention logits
+        # drift as the projection norms are learned.
+        self.use_qk_norm = config.use_qk_norm
+        if self.use_qk_norm:
+            self.q_norm = RMSNorm(
+                hidden_size=self.head_dim,
+                eps=config.rms_norm_eps,
+            )
+            self.k_norm = RMSNorm(
+                hidden_size=self.head_dim,
+                eps=config.rms_norm_eps,
+            )
+
+    def forward(
+        self,
+        hidden_states: Tensor,
+        position_embeddings: tuple[Tensor, Tensor],
+        attention_mask: Tensor | None = None,
+        past_key_values: Cache | None = None,
+        output_attentions: bool = False,
+        **kwargs: Any,
+    ) -> Tensor | tuple[Tensor, Tensor | None]:
+        if hidden_states.ndim != 3:
+            raise ValueError(
+                "hidden_states must have shape "
+                "(batch_size, sequence_length, hidden_size), "
+                f"got shape={tuple(hidden_states.shape)}"
+            )
+
+        batch_size, sequence_length, hidden_size = hidden_states.shape
+        if hidden_size != self.hidden_size:
+            raise ValueError(
+                f"Expected hidden_size={self.hidden_size}, "
+                f"got hidden_size={hidden_size}"
+            )
+        if sequence_length == 0:
+            raise ValueError("sequence_length must be greater than zero")
+
+        cos, sin = position_embeddings
+        query_states, key_states, value_states = self._project_qkv(hidden_states)
+        query_states, key_states = apply_rotary_pos_emb(
+            query=query_states,
+            key=key_states,
+            cos=cos,
+            sin=sin,
+        )
+
+        if past_key_values is not None:
+            # Transformers v5 caches track their own write offset; passing
+            # cache_position here was removed from the library's convention.
+            key_states, value_states = past_key_values.update(
+                key_states,
+                value_states,
+                self.layer_idx,
+            )
+
+        implementation = self.config._attn_implementation or "sdpa"
+        if output_attentions:
+            implementation = "eager"
+
+        attention_interface = ALL_ATTENTION_FUNCTIONS.get_interface(
+            implementation,
+            eager_attention_forward,
+        )
+        attention_output, attention_weights = attention_interface(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            dropout=(self.attention_dropout if self.training else 0.0),
+            scaling=self.scaling,
+            output_attentions=output_attentions,
+            **kwargs,
+        )
+
+        attention_output = attention_output.reshape(
+            batch_size,
+            sequence_length,
+            self.query_size,
+        )
+        attention_output = self.out_proj(attention_output)
+
+        if output_attentions:
+            return attention_output, attention_weights
+        return attention_output
+
+    def _project_qkv(
+        self,
+        hidden_states: Tensor,
+    ) -> tuple[Tensor, Tensor, Tensor]:
+        batch_size, sequence_length, _ = hidden_states.shape
+        qkv_states = self.qkv_proj(hidden_states)
+        query_states, key_states, value_states = qkv_states.split(
+            (
+                self.query_size,
+                self.key_value_size,
+                self.key_value_size,
+            ),
+            dim=-1,
+        )
+
+        query_states = query_states.view(
+            batch_size,
+            sequence_length,
+            self.num_heads,
+            self.head_dim,
+        ).transpose(1, 2)
+        key_states = key_states.view(
+            batch_size,
+            sequence_length,
+            self.num_key_value_heads,
+            self.head_dim,
+        ).transpose(1, 2)
+        value_states = value_states.view(
+            batch_size,
+            sequence_length,
+            self.num_key_value_heads,
+            self.head_dim,
+        ).transpose(1, 2)
+
+        # Applied before RoPE so the rotation acts on unit-scale vectors and
+        # the norm never sees position-dependent structure.
+        if self.use_qk_norm:
+            query_states = self.q_norm(query_states)
+            key_states = self.k_norm(key_states)
+
+        return query_states, key_states, value_states
+
+    def _load_from_state_dict(
+        self,
+        state_dict: dict[str, Tensor],
+        prefix: str,
+        local_metadata: dict[str, Any],
+        strict: bool,
+        missing_keys: list[str],
+        unexpected_keys: list[str],
+        error_msgs: list[str],
+    ) -> None:
+        qkv_key = f"{prefix}qkv_proj.weight"
+        qkv_weight = state_dict.get(qkv_key)
+        expected_shape = tuple(self.qkv_proj.weight.shape)
+        if qkv_weight is not None and tuple(qkv_weight.shape) != expected_shape:
+            error_msgs.append(
+                f"{qkv_key} must use fused [Q, K, V] layout with shape "
+                f"{expected_shape}, got {tuple(qkv_weight.shape)}"
+            )
+
+        super()._load_from_state_dict(
+            state_dict,
+            prefix,
+            local_metadata,
+            strict,
+            missing_keys,
+            unexpected_keys,
+            error_msgs,
+        )
+
+    def extra_repr(self) -> str:
+        return (
+            f"hidden_size={self.hidden_size}, "
+            f"num_heads={self.num_heads}, "
+            f"num_key_value_heads={self.num_key_value_heads}, "
+            f"head_dim={self.head_dim}, "
+            f"attention_dropout={self.attention_dropout}, "
+            f"use_qk_norm={self.use_qk_norm}, "
+            f"layer_idx={self.layer_idx}"
+        )

chat_template.jinja

@@ -0,0 +1,4 @@
+{%- for m in messages %}{{ '<|im_start|>' + m['role'] + '
+' + m['content'] + '<|im_end|>' + '
+' }}{%- endfor %}{%- if add_generation_prompt %}{{ '<|im_start|>assistant
+' }}{%- endif %}

config.json

@@ -0,0 +1,35 @@
+{
+  "architectures": [
+    "NeuronLMForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_neuron_lm.NeuronLMConfig",
+    "AutoModelForCausalLM": "modeling_neuron_lm.NeuronLMForCausalLM"
+  },
+  "bos_token_id": 0,
+  "dtype": "bfloat16",
+  "eos_token_id": 5,
+  "hidden_size": 3584,
+  "initializer_range": 0.02,
+  "intermediate_size": 9728,
+  "is_causal": true,
+  "is_decoder": true,
+  "max_position_embeddings": 4096,
+  "model_type": "neuron_lm",
+  "num_attention_heads": 28,
+  "num_hidden_layers": 36,
+  "num_key_value_heads": 4,
+  "pad_token_id": 3,
+  "residual_dropout": 0.0,
+  "rms_norm_eps": 1e-05,
+  "rope_parameters": {
+    "rope_theta": 500000.0,
+    "rope_type": "default"
+  },
+  "tie_word_embeddings": false,
+  "transformers_version": "5.12.1",
+  "use_cache": false,
+  "use_qk_norm": true,
+  "vocab_size": 48000
+}

configuration_neuron_lm.py

@@ -0,0 +1,288 @@
+from __future__ import annotations
+
+import math
+from copy import deepcopy
+from pathlib import Path
+from typing import Any, ClassVar
+
+import yaml
+from transformers import PreTrainedConfig
+
+# Rotary base frequency. 500000 follows Llama-3 rather than Llama-2's 10000:
+# it trades a little short-range frequency resolution for enough headroom to
+# extend context past the frozen 4K presets later. RoPE scaling is not
+# implemented, so this value is fixed for the lifetime of a pretrained model.
+DEFAULT_ROPE_THETA = 500_000.0
+
+
+class NeuronLMConfig(PreTrainedConfig):
+    """Configuration for the NeuronLM decoder-only language model.
+
+    ``rope_theta`` remains accepted as a compatibility alias, but new
+    configurations serialize rotary settings through Transformers v5's
+    ``rope_parameters`` field.
+    """
+
+    model_type = "neuron_lm"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    # Frozen 4K-context training presets live as YAML files under this
+    # directory (one `<name>.yaml` file per preset), not in this module.
+    # Add or change a preset by editing/adding a YAML file, not this class.
+    PRESETS_DIR: ClassVar[Path] = Path("configs/model")
+
+    def __init__(
+        self,
+        vocab_size: int = 32_000,
+        hidden_size: int = 768,
+        intermediate_size: int = 2_048,
+        num_hidden_layers: int = 12,
+        num_attention_heads: int = 12,
+        num_key_value_heads: int | None = None,
+        max_position_embeddings: int = 2_048,
+        rope_parameters: dict[str, Any] | None = None,
+        rope_theta: float | None = None,
+        rms_norm_eps: float = 1e-5,
+        use_qk_norm: bool = True,
+        attention_dropout: float = 0.0,
+        residual_dropout: float = 0.0,
+        initializer_range: float = 0.02,
+        tie_word_embeddings: bool = True,
+        use_cache: bool = True,
+        # <s>=0, </s>=1, <unk>=2, <pad>=3 is the fixed special-token order
+        # every NeuronLM tokenizer trains with (pretokenization.py's
+        # DEFAULT_SPECIAL_TOKENS). Defaulting here means .generate() stops at
+        # EOS without the caller having to pass it explicitly, and it can
+        # still be overridden for a tokenizer with a different layout.
+        pad_token_id: int | None = 3,
+        bos_token_id: int | None = 0,
+        eos_token_id: int | list[int] | None = 1,
+        **kwargs: Any,
+    ) -> None:
+        if rope_parameters is not None and not isinstance(
+            rope_parameters,
+            dict,
+        ):
+            raise TypeError(
+                "rope_parameters must be a dictionary or None, "
+                f"got {type(rope_parameters).__name__}"
+            )
+
+        resolved_rope_parameters = deepcopy(rope_parameters) or {}
+        if (
+            rope_theta is not None
+            and "rope_theta" in resolved_rope_parameters
+            and resolved_rope_parameters["rope_theta"] != rope_theta
+        ):
+            raise ValueError(
+                "rope_theta and rope_parameters['rope_theta'] disagree: "
+                f"{rope_theta!r} != "
+                f"{resolved_rope_parameters['rope_theta']!r}"
+            )
+        resolved_rope_parameters.setdefault(
+            "rope_theta",
+            DEFAULT_ROPE_THETA if rope_theta is None else rope_theta,
+        )
+        resolved_rope_parameters.setdefault("rope_type", "default")
+
+        self.vocab_size = vocab_size
+        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.num_key_value_heads = (
+            num_attention_heads if num_key_value_heads is None else num_key_value_heads
+        )
+        self.max_position_embeddings = max_position_embeddings
+        self.rope_parameters = resolved_rope_parameters
+        self.rms_norm_eps = rms_norm_eps
+        self.use_qk_norm = use_qk_norm
+        self.attention_dropout = attention_dropout
+        self.residual_dropout = residual_dropout
+        self.initializer_range = initializer_range
+        self.is_decoder = True
+        self.is_encoder_decoder = False
+        self.is_causal = True
+        self.use_cache = use_cache
+
+        self._validate_dimensions()
+
+        kwargs.update(
+            {
+                "pad_token_id": pad_token_id,
+                "bos_token_id": bos_token_id,
+                "eos_token_id": eos_token_id,
+                "tie_word_embeddings": tie_word_embeddings,
+            }
+        )
+        super().__init__(**kwargs)
+
+        self._validate()
+        self.validate_rope()
+
+    @classmethod
+    def available_presets(cls, presets_dir: str | Path | None = None) -> list[str]:
+        """List preset names discoverable as YAML files in ``presets_dir``."""
+
+        directory = Path(presets_dir) if presets_dir is not None else cls.PRESETS_DIR
+        if not directory.is_dir():
+            return []
+        return sorted(path.stem for path in directory.glob("*.yaml"))
+
+    @classmethod
+    def _load_preset(cls, name: str, presets_dir: str | Path | None) -> dict[str, Any]:
+        directory = Path(presets_dir) if presets_dir is not None else cls.PRESETS_DIR
+        preset_path = directory / f"{name}.yaml"
+        try:
+            with preset_path.open("r", encoding="utf-8") as handle:
+                preset = yaml.safe_load(handle)
+        except OSError as error:
+            available = ", ".join(cls.available_presets(directory))
+            raise ValueError(
+                f"Unknown NeuronLM preset {name!r}; available presets: {available}"
+            ) from error
+
+        if not isinstance(preset, dict):
+            raise ValueError(
+                f"preset file {preset_path} must contain a YAML mapping of "
+                "structural fields"
+            )
+        return preset
+
+    @classmethod
+    def from_preset(
+        cls,
+        name: str,
+        *,
+        vocab_size: int = 32_000,
+        presets_dir: str | Path | None = None,
+        **overrides: Any,
+    ) -> NeuronLMConfig:
+        """Construct one of the frozen 4K-context training presets.
+
+        Presets are loaded from YAML files under ``presets_dir`` (defaults
+        to ``cls.PRESETS_DIR``), one file per preset named ``<name>.yaml``.
+        """
+
+        preset = deepcopy(cls._load_preset(name, presets_dir))
+
+        structural_fields = set(preset)
+        conflicting = structural_fields.intersection(overrides)
+        if conflicting:
+            names = ", ".join(sorted(conflicting))
+            raise ValueError(
+                f"Preset {name!r} has frozen structural fields and cannot "
+                f"override: {names}"
+            )
+
+        return cls(vocab_size=vocab_size, **preset, **overrides)
+
+    def _validate(self) -> None:
+        self._validate_dimensions()
+
+        if self.head_dim % 2 != 0:
+            raise ValueError(
+                f"RoPE requires an even head dimension, got head_dim={self.head_dim}"
+            )
+
+        if self.rope_parameters.get("rope_type") != "default":
+            raise ValueError(
+                "NeuronLM currently supports only default RoPE; context "
+                "extrapolation methods are intentionally deferred"
+            )
+
+        if not _is_positive_finite_number(self.rope_theta):
+            raise ValueError(f"rope_theta must be positive, got {self.rope_theta}")
+
+        if not _is_positive_finite_number(self.rms_norm_eps):
+            raise ValueError(f"rms_norm_eps must be positive, got {self.rms_norm_eps}")
+
+        if type(self.use_qk_norm) is not bool:
+            raise ValueError(f"use_qk_norm must be a boolean, got {self.use_qk_norm!r}")
+
+        for name, value in {
+            "attention_dropout": self.attention_dropout,
+            "residual_dropout": self.residual_dropout,
+        }.items():
+            if (
+                isinstance(value, bool)
+                or not isinstance(value, (int, float))
+                or not math.isfinite(float(value))
+                or not 0.0 <= value < 1.0
+            ):
+                raise ValueError(f"{name} must be in [0, 1), got {value}")
+
+        if not _is_positive_finite_number(self.initializer_range):
+            raise ValueError(
+                f"initializer_range must be positive, got {self.initializer_range}"
+            )
+
+    def _validate_dimensions(self) -> None:
+        positive_int_fields = {
+            "vocab_size": self.vocab_size,
+            "hidden_size": self.hidden_size,
+            "intermediate_size": self.intermediate_size,
+            "num_hidden_layers": self.num_hidden_layers,
+            "num_attention_heads": self.num_attention_heads,
+            "num_key_value_heads": self.num_key_value_heads,
+            "max_position_embeddings": self.max_position_embeddings,
+        }
+
+        for name, value in positive_int_fields.items():
+            if type(value) is not int or value <= 0:
+                raise ValueError(f"{name} must be a positive integer, got {value!r}")
+
+        if self.hidden_size % self.num_attention_heads != 0:
+            raise ValueError(
+                f"hidden_size={self.hidden_size} must be divisible by "
+                f"num_attention_heads={self.num_attention_heads}"
+            )
+
+        if self.num_attention_heads % self.num_key_value_heads != 0:
+            raise ValueError(
+                "num_attention_heads must be divisible by "
+                "num_key_value_heads, got "
+                f"{self.num_attention_heads} and "
+                f"{self.num_key_value_heads}"
+            )
+
+    @property
+    def head_dim(self) -> int:
+        return self.hidden_size // self.num_attention_heads
+
+    @property
+    def qkv_projection_size(self) -> int:
+        return (self.num_attention_heads + 2 * self.num_key_value_heads) * self.head_dim
+
+    @property
+    def rope_theta(self) -> float:
+        return float(self.rope_parameters["rope_theta"])
+
+    @property
+    def d_model(self) -> int:
+        return self.hidden_size
+
+    @property
+    def d_ff(self) -> int:
+        return self.intermediate_size
+
+    @property
+    def num_layers(self) -> int:
+        return self.num_hidden_layers
+
+    @property
+    def num_heads(self) -> int:
+        return self.num_attention_heads
+
+    @property
+    def context_length(self) -> int:
+        return self.max_position_embeddings
+
+
+def _is_positive_finite_number(value: Any) -> bool:
+    return (
+        not isinstance(value, bool)
+        and isinstance(value, (int, float))
+        and math.isfinite(float(value))
+        and value > 0
+    )

generation_config.json

@@ -0,0 +1,11 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 0,
+  "do_sample": true,
+  "eos_token_id": 5,
+  "output_attentions": false,
+  "output_hidden_states": false,
+  "pad_token_id": 3,
+  "transformers_version": "5.12.1",
+  "use_cache": true
+}

layers.py

@@ -0,0 +1,88 @@
+from __future__ import annotations
+
+import torch
+import torch.nn.functional as F
+from torch import Tensor, nn
+
+__all__ = [
+    "RMSNorm",
+    "SwiGLU",
+]
+
+
+class RMSNorm(nn.RMSNorm):
+    def __init__(
+        self,
+        hidden_size: int,
+        eps: float = 1e-5,
+        *,
+        device: torch.device | str | None = None,
+        dtype: torch.dtype | None = None,
+    ) -> None:
+        if hidden_size <= 0:
+            raise ValueError(f"hidden_size must be positive, got {hidden_size}")
+
+        if eps <= 0.0:
+            raise ValueError(f"eps must be positive, got {eps}")
+
+        super().__init__(
+            normalized_shape=hidden_size,
+            eps=eps,
+            elementwise_affine=True,
+            device=device,
+            dtype=dtype,
+        )
+
+        self.hidden_size = hidden_size
+
+
+class SwiGLU(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        *,
+        device: torch.device | str | None = None,
+        dtype: torch.dtype | None = None,
+    ) -> None:
+        super().__init__()
+
+        if hidden_size <= 0:
+            raise ValueError(f"hidden_size must be positive, got {hidden_size}")
+
+        if intermediate_size <= 0:
+            raise ValueError(
+                f"intermediate_size must be positive, got {intermediate_size}"
+            )
+
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+
+        self.gate_up_proj = nn.Linear(
+            in_features=hidden_size,
+            out_features=2 * intermediate_size,
+            bias=False,
+            device=device,
+            dtype=dtype,
+        )
+
+        self.down_proj = nn.Linear(
+            in_features=intermediate_size,
+            out_features=hidden_size,
+            bias=False,
+            device=device,
+            dtype=dtype,
+        )
+
+    def forward(self, hidden_states: Tensor) -> Tensor:
+        gate, up = self.gate_up_proj(hidden_states).chunk(2, dim=-1)
+
+        hidden_states = F.silu(gate) * up
+        return self.down_proj(hidden_states)
+
+    def extra_repr(self) -> str:
+        return (
+            f"hidden_size={self.hidden_size}, "
+            f"intermediate_size={self.intermediate_size}, "
+            "bias=False"
+        )

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ac18b7d0f870b7d82c774b9cca60e246031683ade0c54d8dac8b741bbbd51d6
+size 11022175080

modeling_neuron_lm.py

@@ -0,0 +1,644 @@
+from __future__ import annotations
+
+import math
+from copy import copy
+from typing import Any, cast
+
+import torch
+from torch import Tensor, nn
+from torch.utils.checkpoint import checkpoint
+from transformers import Cache, DynamicCache, PreTrainedModel
+from transformers.generation.utils import GenerationMixin
+from transformers.masking_utils import create_causal_mask
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+)
+
+from .attention import NeuronLMAttention
+from .configuration_neuron_lm import NeuronLMConfig
+from .layers import RMSNorm, SwiGLU
+from .rotary import RotaryEmbedding
+
+__all__ = [
+    "NeuronLMDecoderLayer",
+    "NeuronLMPreTrainedModel",
+    "NeuronLMModel",
+    "NeuronLMForCausalLM",
+]
+
+# Marks the linear projection that writes a residual branch back into the
+# residual stream. ``_init_weights`` scales these down by
+# ``1 / sqrt(2 * num_hidden_layers)`` so residual-stream variance stays
+# roughly constant with depth at initialization (GPT-2 / OLMo convention).
+# Set through ``setattr`` because ``nn.Module.__setattr__`` is typed for
+# parameters, buffers, and submodules only.
+RESIDUAL_PROJECTION_FLAG = "_neuron_lm_residual_projection"
+
+
+def _cache_seq_length(cache: Cache | None, layer_idx: int = 0) -> int:
+    if cache is None:
+        return 0
+    length = cache.get_seq_length(layer_idx)
+    if isinstance(length, Tensor):
+        # ``.item()`` is a graph break under torch.compile. Callers only reach
+        # this path when they did not supply position_ids, and generation
+        # always supplies them.
+        length = length.item()
+    return int(length)
+
+
+class NeuronLMDecoderLayer(nn.Module):
+    def __init__(
+        self,
+        config: NeuronLMConfig,
+        layer_idx: int,
+    ) -> None:
+        super().__init__()
+
+        if type(layer_idx) is not int or layer_idx < 0:
+            raise ValueError(
+                f"layer_idx must be a non-negative integer, got {layer_idx!r}"
+            )
+
+        self.hidden_size = config.hidden_size
+        self.layer_idx = layer_idx
+
+        self.input_layernorm = RMSNorm(
+            hidden_size=config.hidden_size,
+            eps=config.rms_norm_eps,
+        )
+
+        self.self_attn = NeuronLMAttention(
+            config,
+            layer_idx=layer_idx,
+        )
+
+        self.post_attention_layernorm = RMSNorm(
+            hidden_size=config.hidden_size,
+            eps=config.rms_norm_eps,
+        )
+
+        self.mlp = SwiGLU(
+            hidden_size=config.hidden_size,
+            intermediate_size=config.intermediate_size,
+        )
+
+        self.residual_dropout = nn.Dropout(
+            p=config.residual_dropout,
+        )
+
+        # Both branches of this layer write through these two projections.
+        setattr(self.self_attn.out_proj, RESIDUAL_PROJECTION_FLAG, True)
+        setattr(self.mlp.down_proj, RESIDUAL_PROJECTION_FLAG, True)
+
+    def forward(
+        self,
+        hidden_states: Tensor,
+        position_embeddings: tuple[Tensor, Tensor],
+        attention_mask: Tensor | None = None,
+        past_key_values: Cache | None = None,
+        output_attentions: bool = False,
+    ) -> Tensor | tuple[Tensor, Tensor | None]:
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+        attention_outputs = self.self_attn(
+            hidden_states=hidden_states,
+            position_embeddings=position_embeddings,
+            attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            output_attentions=output_attentions,
+        )
+        if output_attentions:
+            hidden_states, attention_weights = attention_outputs
+        else:
+            hidden_states = attention_outputs
+            attention_weights = None
+        hidden_states = residual + self.residual_dropout(hidden_states)
+
+        residual = hidden_states
+
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + self.residual_dropout(hidden_states)
+
+        if output_attentions:
+            return hidden_states, attention_weights
+        return hidden_states
+
+
+class NeuronLMPreTrainedModel(PreTrainedModel):
+    config_class = NeuronLMConfig
+    base_model_prefix = "model"
+
+    supports_gradient_checkpointing = True
+
+    _no_split_modules = ["NeuronLMDecoderLayer"]
+
+    # Backends verified against the SDPA reference in tests/test_attention.py.
+    # `_supports_flash_attn` stays unset: flash-attn is not installed here, so
+    # the claim cannot be tested, and SDPA already dispatches flash kernels on
+    # recent hardware. FlexAttention is what intra-document masking compiles
+    # its BlockMask through.
+    _supports_sdpa = True
+    _supports_flex_attn = True
+
+    # The forward is free of data-dependent control flow, so Transformers may
+    # use its compiled generation path. Regression coverage:
+    # tests/test_modeling.py::test_forward_compiles_as_a_full_graph.
+    _can_compile_fullgraph = True
+
+    # _tp_plan is intentionally unset. The fused qkv_proj packs three blocks
+    # whose sizes follow the GQA head counts (num_attention_heads,
+    # num_key_value_heads, num_key_value_heads), while Transformers'
+    # "packed_colwise" style assumes two equally sized blocks -- it would cut
+    # through the K block. Supporting tensor parallelism here needs both a
+    # custom sharding style and a _project_qkv that splits on per-rank head
+    # counts.
+    #
+    # That work is not on the critical path: FSDP2 (configs/accelerate/
+    # fsdp2.yaml) shards an 8B AdamW run to roughly 30 GiB per GPU across
+    # four devices, so memory is not the binding constraint at the sizes this
+    # model targets. Revisit if serving latency or a much larger model makes
+    # tensor parallelism necessary; TrainingArguments.parallelism_config is
+    # the entry point.
+
+    def residual_initializer_std(self) -> float:
+        """Initialization std for projections feeding the residual stream.
+
+        Scaling by ``1 / sqrt(2 * num_hidden_layers)`` keeps the variance of
+        the residual stream from growing with depth. There are two residual
+        branches per decoder layer, hence the factor of two.
+        """
+
+        depth_scale = math.sqrt(2.0 * self.config.num_hidden_layers)
+        return self.config.initializer_range / depth_scale
+
+    def _init_weights(self, module: nn.Module) -> None:
+
+        if isinstance(module, nn.Linear):
+            if getattr(module, RESIDUAL_PROJECTION_FLAG, False):
+                std = self.residual_initializer_std()
+            else:
+                std = self.config.initializer_range
+
+            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=self.config.initializer_range,
+            )
+
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+        elif isinstance(module, nn.RMSNorm):
+            if module.elementwise_affine:
+                module.weight.data.fill_(1.0)
+
+        elif isinstance(module, RotaryEmbedding):
+            module.reset_parameters()
+
+
+class NeuronLMModel(NeuronLMPreTrainedModel):
+    def __init__(self, config: NeuronLMConfig) -> None:
+        super().__init__(config)
+
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(
+            num_embeddings=config.vocab_size,
+            embedding_dim=config.hidden_size,
+            padding_idx=config.pad_token_id,
+        )
+
+        self.layers = nn.ModuleList(
+            [
+                NeuronLMDecoderLayer(
+                    config=config,
+                    layer_idx=layer_idx,
+                )
+                for layer_idx in range(config.num_hidden_layers)
+            ]
+        )
+
+        self.norm = RMSNorm(
+            hidden_size=config.hidden_size,
+            eps=config.rms_norm_eps,
+        )
+
+        # RoPE frequencies are computed once per model forward and shared by
+        # all decoder layers.
+        self.rotary_emb = RotaryEmbedding(
+            head_dim=config.head_dim,
+            base=config.rope_theta,
+        )
+
+        # PreTrainedModel.gradient_checkpointing_enable() updates this flag
+        # and assigns self._gradient_checkpointing_func.
+        self.gradient_checkpointing = False
+
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Embedding:
+        return self.embed_tokens
+
+    def set_input_embeddings(
+        self,
+        value: nn.Embedding,
+    ) -> None:
+        self.embed_tokens = value
+
+    def forward(
+        self,
+        input_ids: Tensor | None = None,
+        attention_mask: Tensor | None = None,
+        position_ids: Tensor | None = None,
+        inputs_embeds: Tensor | None = None,
+        past_key_values: Cache | None = None,
+        use_cache: bool | None = None,
+        output_attentions: bool | None = None,
+        output_hidden_states: bool | None = None,
+        return_dict: bool | None = None,
+        **kwargs: Any,
+    ) -> BaseModelOutputWithPast | tuple[Tensor, ...]:
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+        return_dict = (
+            return_dict if return_dict is not None else self.config.return_dict
+        )
+        use_cache = (
+            use_cache
+            if use_cache is not None
+            else (self.config.use_cache or past_key_values is not None)
+        )
+
+        if kwargs:
+            unsupported = ", ".join(sorted(kwargs))
+            raise TypeError(f"Unsupported model forward arguments: {unsupported}")
+
+        if past_key_values is not None and not isinstance(
+            past_key_values,
+            Cache,
+        ):
+            raise TypeError(
+                "past_key_values must be a Hugging Face Cache instance; "
+                "legacy tuple caches are not supported"
+            )
+
+        # Cache mutation is incompatible with recomputation during backward.
+        # This mirrors the behavior of the current Transformers decoder
+        # layers while keeping the public forward API convenient.
+        if self.gradient_checkpointing and self.training:
+            use_cache = False
+            past_key_values = None
+
+        if use_cache:
+            if past_key_values is None:
+                past_key_values = DynamicCache(config=self.config)
+        elif past_key_values is not None:
+            raise ValueError("past_key_values can only be used when use_cache=True")
+
+        if (input_ids is None) == (inputs_embeds is None):
+            raise ValueError("Specify exactly one of input_ids or inputs_embeds")
+
+        if input_ids is not None:
+            if input_ids.ndim != 2:
+                raise ValueError(
+                    "input_ids must have shape "
+                    "(batch_size, sequence_length), "
+                    f"got shape={tuple(input_ids.shape)}"
+                )
+
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        assert inputs_embeds is not None
+
+        if inputs_embeds.ndim != 3:
+            raise ValueError(
+                "inputs_embeds must have shape "
+                "(batch_size, sequence_length, hidden_size), "
+                f"got shape={tuple(inputs_embeds.shape)}"
+            )
+
+        batch_size, sequence_length, hidden_size = inputs_embeds.shape
+
+        if hidden_size != self.config.hidden_size:
+            raise ValueError(
+                f"Expected hidden_size={self.config.hidden_size}, "
+                f"got hidden_size={hidden_size}"
+            )
+
+        if sequence_length == 0:
+            raise ValueError("sequence_length must be greater than zero")
+
+        past_key_length = (
+            _cache_seq_length(past_key_values) if past_key_values is not None else 0
+        )
+        hidden_states = inputs_embeds
+
+        if position_ids is None:
+            position_ids = torch.arange(
+                past_key_length,
+                past_key_length + sequence_length,
+                dtype=torch.long,
+                device=hidden_states.device,
+            ).unsqueeze(0)
+        else:
+            if position_ids.ndim not in {1, 2}:
+                raise ValueError(
+                    "position_ids must have shape (sequence_length,) or "
+                    "(batch_size, sequence_length), "
+                    f"got shape={tuple(position_ids.shape)}"
+                )
+            if position_ids.shape[-1] != sequence_length:
+                raise ValueError(
+                    "The final position_ids dimension must equal the "
+                    f"sequence length {sequence_length}, got "
+                    f"{position_ids.shape[-1]}"
+                )
+            position_ids = position_ids.to(
+                device=hidden_states.device,
+                dtype=torch.long,
+            )
+            if position_ids.ndim == 1:
+                position_ids = position_ids.unsqueeze(0)
+
+        # Transformers derives packed-document boundaries from gaps in
+        # position_ids, and that detection requires a 2D tensor.
+        # See create_causal_mask / find_packed_sequence_indices.
+
+        # Reading position_ids.max() is data-dependent control flow, which
+        # torch.compile cannot trace in a full graph. The bound is a static
+        # property of the config, so the eager check is sufficient: any shape
+        # that would trip it also trips it before compilation warms up.
+        if (
+            not torch.compiler.is_compiling()
+            and position_ids.numel() > 0
+            and position_ids.max() >= self.config.max_position_embeddings
+        ):
+            raise ValueError(
+                "position_ids contain a position at or beyond "
+                f"max_position_embeddings={self.config.max_position_embeddings}"
+            )
+
+        position_embeddings = self.rotary_emb(
+            hidden_states,
+            position_ids=position_ids,
+        )
+
+        mask_config = self.config
+        if output_attentions and self.config._attn_implementation != "eager":
+            mask_config = copy(self.config)
+            mask_config._attn_implementation = "eager"
+
+        causal_attention_mask = create_causal_mask(
+            config=mask_config,
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            position_ids=position_ids,
+        )
+
+        all_hidden_states: tuple[Tensor, ...] | None = (
+            () if output_hidden_states else None
+        )
+        all_self_attentions: tuple[Tensor, ...] | None = (
+            () if output_attentions else None
+        )
+
+        for decoder_layer in self.layers:
+            decoder_layer = cast(NeuronLMDecoderLayer, decoder_layer)
+            if all_hidden_states is not None:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+
+                def custom_forward(
+                    states: Tensor,
+                    layer: NeuronLMDecoderLayer = decoder_layer,
+                ) -> Tensor | tuple[Tensor, Tensor | None]:
+                    return layer(
+                        hidden_states=states,
+                        position_embeddings=position_embeddings,
+                        attention_mask=causal_attention_mask,
+                        past_key_values=None,
+                        output_attentions=output_attentions,
+                    )
+
+                checkpointing_function = getattr(
+                    self,
+                    "_gradient_checkpointing_func",
+                    None,
+                )
+
+                if checkpointing_function is None:
+                    layer_outputs = checkpoint(
+                        custom_forward,
+                        hidden_states,
+                        use_reentrant=False,
+                    )
+                else:
+                    layer_outputs = checkpointing_function(
+                        custom_forward,
+                        hidden_states,
+                    )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states=hidden_states,
+                    position_embeddings=position_embeddings,
+                    attention_mask=causal_attention_mask,
+                    past_key_values=past_key_values,
+                    output_attentions=output_attentions,
+                )
+
+            if output_attentions:
+                hidden_states, attention_weights = layer_outputs
+                assert all_self_attentions is not None
+                assert attention_weights is not None
+                all_self_attentions += (attention_weights,)
+            else:
+                hidden_states = layer_outputs
+
+        hidden_states = self.norm(hidden_states)
+
+        if all_hidden_states is not None:
+            all_hidden_states += (hidden_states,)
+
+        if not return_dict:
+            outputs: tuple[Any, ...] = (hidden_states,)
+
+            if use_cache:
+                outputs += (past_key_values,)
+
+            if output_hidden_states:
+                outputs += (all_hidden_states,)
+
+            if output_attentions:
+                outputs += (all_self_attentions,)
+
+            return outputs
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values if use_cache else None,
+            hidden_states=cast(Any, all_hidden_states),
+            attentions=cast(Any, all_self_attentions),
+        )
+
+
+class NeuronLMForCausalLM(
+    NeuronLMPreTrainedModel,
+    GenerationMixin,
+):
+    """NeuronLM decoder with a causal language-modeling head."""
+
+    _tied_weights_keys = {
+        "lm_head.weight": "model.embed_tokens.weight",
+    }
+
+    def __init__(self, config: NeuronLMConfig) -> None:
+        super().__init__(config)
+
+        self.model = NeuronLMModel(config)
+        self.vocab_size = config.vocab_size
+
+        self.lm_head = nn.Linear(
+            in_features=config.hidden_size,
+            out_features=config.vocab_size,
+            bias=False,
+        )
+
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Embedding:
+        return self.model.embed_tokens
+
+    def set_input_embeddings(
+        self,
+        value: nn.Embedding,
+    ) -> None:
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self) -> nn.Linear:
+        return self.lm_head
+
+    def set_output_embeddings(
+        self,
+        value: nn.Linear,
+    ) -> None:
+        self.lm_head = value
+
+    def get_decoder(self) -> NeuronLMModel:
+        return self.model
+
+    def set_decoder(
+        self,
+        decoder: NeuronLMModel,
+    ) -> None:
+        self.model = decoder
+
+    def forward(
+        self,
+        input_ids: Tensor | None = None,
+        attention_mask: Tensor | None = None,
+        position_ids: Tensor | None = None,
+        inputs_embeds: Tensor | None = None,
+        labels: Tensor | None = None,
+        past_key_values: Cache | None = None,
+        use_cache: bool | None = None,
+        output_attentions: bool | None = None,
+        output_hidden_states: bool | None = None,
+        return_dict: bool | None = None,
+        num_items_in_batch: Tensor | int | None = None,
+        **kwargs: Any,
+    ) -> CausalLMOutputWithPast | tuple[Tensor, ...]:
+        return_dict = (
+            return_dict if return_dict is not None else self.config.return_dict
+        )
+
+        model_outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            inputs_embeds=inputs_embeds,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            **kwargs,
+        )
+
+        if return_dict:
+            hidden_states = model_outputs.last_hidden_state
+        else:
+            hidden_states = model_outputs[0]
+
+        logits = self.lm_head(hidden_states)
+
+        loss: Tensor | None = None
+
+        if labels is not None:
+            if labels.ndim != 2:
+                raise ValueError(
+                    "labels must have shape "
+                    "(batch_size, sequence_length), "
+                    f"got shape={tuple(labels.shape)}"
+                )
+
+            expected_shape = hidden_states.shape[:2]
+
+            if tuple(labels.shape) != tuple(expected_shape):
+                raise ValueError(
+                    f"labels must have shape {tuple(expected_shape)}, "
+                    f"got {tuple(labels.shape)}"
+                )
+
+            if labels.shape[1] < 2:
+                raise ValueError(
+                    "At least two sequence positions are required "
+                    "to compute causal language-modeling loss"
+                )
+
+            labels = labels.to(device=logits.device)
+
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                vocab_size=self.config.vocab_size,
+                num_items_in_batch=num_items_in_batch,
+            )
+
+        if not return_dict:
+            output = (logits,) + model_outputs[1:]
+
+            if loss is not None:
+                return (loss,) + output
+
+            return output
+
+        return CausalLMOutputWithPast(
+            loss=cast(Any, loss),
+            logits=logits,
+            past_key_values=model_outputs.past_key_values,
+            hidden_states=model_outputs.hidden_states,
+            attentions=model_outputs.attentions,
+        )

rotary.py

@@ -0,0 +1,306 @@
+"""Adjacent-pair (GPT-J style) rotary position embeddings.
+
+Convention, which matters when exporting a trained checkpoint: this module
+rotates *adjacent* channel pairs ``(x0, x1), (x2, x3), ...``. Llama and most
+Hugging Face models instead rotate *half-split* pairs ``(x0, x_{d/2}), ...``
+("NeoX style"). The two are related by a permutation of the query/key rows,
+so a checkpoint trained here is NOT drop-in loadable as a Llama checkpoint
+without permuting ``qkv_proj``.
+
+Both conventions are first-class in the common inference runtimes -- select
+GPT-J/``NORM``-style rotary rather than ``NEOX`` when converting. Concretely:
+``llama.cpp`` ``rope_type=NORM``, vLLM ``is_neox_style=False``.
+
+``cos``/``sin`` here have shape ``(..., sequence_length, head_dim / 2)``,
+half the width of the Hugging Face convention, because adjacent-pair rotation
+needs one angle per pair rather than a duplicated pair of angles. That makes
+this form measurably cheaper than the half-split ``rotate_half`` formulation,
+which needs full-width tables and a concatenation.
+
+Regression coverage for the convention itself lives in
+``tests/test_rotary.py::manual_adjacent_pair_rotation``.
+"""
+
+from __future__ import annotations
+
+import math
+
+import torch
+from torch import Tensor, nn
+
+__all__ = [
+    "RotaryEmbedding",
+    "apply_rotary_pos_emb",
+]
+
+
+_INTEGER_DTYPES = {
+    torch.uint8,
+    torch.int8,
+    torch.int16,
+    torch.int32,
+    torch.int64,
+}
+
+
+class RotaryEmbedding(nn.Module):
+    def __init__(
+        self,
+        head_dim: int,
+        base: float = 10_000.0,
+        *,
+        device: torch.device | str | None = None,
+    ) -> None:
+        super().__init__()
+
+        if type(head_dim) is not int or head_dim <= 0:
+            raise ValueError(f"head_dim must be a positive integer, got {head_dim!r}")
+
+        if head_dim % 2 != 0:
+            raise ValueError(f"head_dim must be even, got head_dim={head_dim}")
+
+        if (
+            isinstance(base, bool)
+            or not isinstance(base, (int, float))
+            or not math.isfinite(float(base))
+            or base <= 0.0
+        ):
+            raise ValueError(f"base must be a positive finite number, got {base!r}")
+
+        self.head_dim = head_dim
+        self.base = float(base)
+
+        self.register_buffer(
+            "inv_freq",
+            torch.empty(
+                head_dim // 2,
+                dtype=torch.float32,
+                device=device,
+            ),
+            persistent=False,
+        )
+        self.reset_parameters()
+
+    def reset_parameters(self) -> None:
+        """Reconstruct inverse frequencies on the buffer's current device."""
+
+        frequency_indices = torch.arange(
+            start=0,
+            end=self.head_dim,
+            step=2,
+            dtype=torch.float32,
+            device=self.inv_freq.device,
+        )
+
+        inv_freq = self.base ** (-frequency_indices / self.head_dim)
+
+        # Assignment preserves the registered, non-persistent buffer while
+        # also replacing storage allocated by Transformers' meta-device
+        # loading path.
+        self.inv_freq = inv_freq
+
+    @torch.no_grad()
+    def forward(
+        self,
+        hidden_states: Tensor,
+        position_ids: Tensor | None = None,
+    ) -> tuple[Tensor, Tensor]:
+
+        if hidden_states.ndim < 2:
+            raise ValueError(
+                "hidden_states must have at least two dimensions, "
+                f"got shape={tuple(hidden_states.shape)}"
+            )
+
+        if not hidden_states.is_floating_point():
+            raise TypeError(
+                "hidden_states must be a floating-point tensor, "
+                f"got dtype={hidden_states.dtype}"
+            )
+
+        sequence_length = hidden_states.shape[-2]
+
+        if position_ids is None:
+            position_ids = torch.arange(
+                sequence_length,
+                device=hidden_states.device,
+                dtype=torch.long,
+            )
+        else:
+            if position_ids.ndim not in {1, 2}:
+                raise ValueError(
+                    "position_ids must have shape "
+                    "(sequence_length,) or "
+                    "(batch_size, sequence_length), "
+                    f"got shape={tuple(position_ids.shape)}"
+                )
+
+            if position_ids.shape[-1] != sequence_length:
+                raise ValueError(
+                    "The final position_ids dimension must equal the "
+                    f"sequence length {sequence_length}, "
+                    f"got {position_ids.shape[-1]}"
+                )
+
+            if position_ids.dtype not in _INTEGER_DTYPES:
+                raise TypeError(
+                    "position_ids must contain integers, "
+                    f"got dtype={position_ids.dtype}"
+                )
+
+            position_ids = position_ids.to(
+                device=hidden_states.device,
+            )
+
+        # Compute frequencies in float32 even when the model is running in
+        # float16 or bfloat16. Cast only the final cosine/sine tensors.
+        inv_freq = self.inv_freq.to(
+            device=hidden_states.device,
+            dtype=torch.float32,
+        )
+
+        positions = position_ids.to(dtype=torch.float32)
+
+        angles = positions.unsqueeze(-1) * inv_freq
+
+        cos = angles.cos()
+        sin = angles.sin()
+
+        return (
+            cos.to(dtype=hidden_states.dtype),
+            sin.to(dtype=hidden_states.dtype),
+        )
+
+    def extra_repr(self) -> str:
+        return f"head_dim={self.head_dim}, base={self.base}"
+
+
+def _reshape_frequencies_for_broadcast(
+    frequencies: Tensor,
+    target: Tensor,
+) -> Tensor:
+
+    extra_dimensions = target.ndim - frequencies.ndim
+
+    if extra_dimensions < 0:
+        raise ValueError(
+            "Rotary frequencies have too many dimensions for the target: "
+            f"frequencies.ndim={frequencies.ndim}, "
+            f"target.ndim={target.ndim}"
+        )
+
+    broadcast_shape = (
+        *frequencies.shape[:-2],
+        *((1,) * extra_dimensions),
+        *frequencies.shape[-2:],
+    )
+
+    return frequencies.reshape(broadcast_shape)
+
+
+def _apply_rotary(
+    hidden_states: Tensor,
+    cos: Tensor,
+    sin: Tensor,
+) -> Tensor:
+
+    if hidden_states.shape[-1] % 2 != 0:
+        raise ValueError(
+            f"The final hidden dimension must be even, got {hidden_states.shape[-1]}"
+        )
+
+    even_states = hidden_states[..., 0::2]
+    odd_states = hidden_states[..., 1::2]
+
+    cos = _reshape_frequencies_for_broadcast(
+        cos,
+        even_states,
+    )
+    sin = _reshape_frequencies_for_broadcast(
+        sin,
+        even_states,
+    )
+
+    rotated_even = even_states * cos - odd_states * sin
+    rotated_odd = even_states * sin + odd_states * cos
+
+    return torch.stack(
+        (rotated_even, rotated_odd),
+        dim=-1,
+    ).flatten(start_dim=-2)
+
+
+def apply_rotary_pos_emb(
+    query: Tensor,
+    key: Tensor,
+    cos: Tensor,
+    sin: Tensor,
+) -> tuple[Tensor, Tensor]:
+
+    if query.ndim < 2 or key.ndim < 2:
+        raise ValueError("query and key must each have at least two dimensions")
+
+    if query.shape[-2] != key.shape[-2]:
+        raise ValueError(
+            "query and key sequence lengths must match, "
+            f"got {query.shape[-2]} and {key.shape[-2]}"
+        )
+
+    if query.shape[-1] != key.shape[-1]:
+        raise ValueError(
+            "query and key head dimensions must match, "
+            f"got {query.shape[-1]} and {key.shape[-1]}"
+        )
+
+    if query.shape[-1] % 2 != 0:
+        raise ValueError(
+            f"The query/key head dimension must be even, got {query.shape[-1]}"
+        )
+
+    if query.device != key.device:
+        raise ValueError(
+            "query and key must be on the same device, "
+            f"got {query.device} and {key.device}"
+        )
+
+    if query.dtype != key.dtype:
+        raise ValueError(
+            f"query and key must have the same dtype, got {query.dtype} and {key.dtype}"
+        )
+
+    if cos.shape != sin.shape:
+        raise ValueError(
+            "cos and sin must have identical shapes, "
+            f"got {tuple(cos.shape)} and {tuple(sin.shape)}"
+        )
+
+    expected_frequency_shape = (
+        query.shape[-2],
+        query.shape[-1] // 2,
+    )
+
+    if cos.shape[-2:] != expected_frequency_shape:
+        raise ValueError(
+            "The final cosine/sine dimensions must be "
+            "(sequence_length, head_dim / 2), "
+            f"expected {expected_frequency_shape}, "
+            f"got {tuple(cos.shape[-2:])}"
+        )
+
+    if cos.device != query.device or sin.device != query.device:
+        raise ValueError("query, key, cos, and sin must be on the same device")
+
+    # PATCHED (see scripts/prepare_neuronai_5b_base.py): align cos/sin with
+    # the query dtype instead of rejecting the pair. Under mixed precision the
+    # qkv projections emit bf16 while hidden_states -- and therefore cos/sin --
+    # stay fp32, which is normal and which upstream HF models handle by
+    # implicit type promotion.
+    if cos.dtype != query.dtype:
+        cos = cos.to(dtype=query.dtype)
+    if sin.dtype != query.dtype:
+        sin = sin.to(dtype=query.dtype)
+
+    return (
+        _apply_rotary(query, sin=sin, cos=cos),
+        _apply_rotary(key, sin=sin, cos=cos),
+    )

tokenizer_config.json

@@ -0,0 +1,12 @@
+{
+  "backend": "tokenizers",
+  "bos_token": "<s>",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "is_local": true,
+  "local_files_only": true,
+  "model_max_length": 4096,
+  "pad_token": "<pad>",
+  "tokenizer_class": "TokenizersBackend",
+  "unk_token": "<unk>"
+}