checkpoint_folder_full/modeling_nemotron_dense.py

"""HuggingFace custom modeling for Nemotron-Dense (Cosmos 2B dense) checkpoints.

Loaded via `AutoModelForCausalLM.from_pretrained(..., trust_remote_code=True)`
using the `auto_map` field in `config.json`. RMSNorm + squared_relu MLP + GQA.
Uses transformers >=4.38 DynamicCache.update() API.

NemotronDenseConfig is a standalone PretrainedConfig (not a NemotronConfig
subclass) for version stability; recent transformers versions migrated
`rope_theta` into a `rope_parameters` dict on NemotronConfig, which breaks
direct attribute access. We follow the modern convention and use
`rope_parameters` exclusively; both the HF modeling code below and the vLLM
plugin read RoPE settings from this dict.
"""

import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import PretrainedConfig, PreTrainedModel
from transformers.activations import ACT2FN
from transformers.cache_utils import DynamicCache
from transformers.generation import GenerationMixin
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast


class NemotronDenseConfig(PretrainedConfig):
    model_type = "nemotron_dense"

    def __init__(
        self,
        vocab_size=131072,
        hidden_size=2048,
        intermediate_size=9216,
        num_hidden_layers=28,
        num_attention_heads=16,
        head_dim=128,
        num_key_value_heads=8,
        hidden_act="relu2",
        max_position_embeddings=131072,
        norm_eps=1e-5,
        rope_parameters=None,
        tie_word_embeddings=False,
        **kwargs,
    ):
        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.head_dim = head_dim
        self.num_key_value_heads = num_key_value_heads
        self.hidden_act = hidden_act
        self.max_position_embeddings = max_position_embeddings
        self.norm_eps = norm_eps
        self.rope_parameters = rope_parameters or {
            "rope_theta": 100000000.0,
            "partial_rotary_factor": 1.0,
        }
        super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)


def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
    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 NemotronDenseRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-5):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, hidden_states):
        return F.rms_norm(hidden_states, self.weight.shape, self.weight, self.variance_epsilon)


class NemotronDenseRotaryEmbedding(nn.Module):
    def __init__(self, dim, max_position_embeddings=131072, base=100000000.0, device=None):
        super().__init__()
        self.dim = dim
        self.max_position_embeddings = max_position_embeddings
        self.base = base
        inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float().to(device) / self.dim))
        self.register_buffer("inv_freq", inv_freq, persistent=False)

    def forward(self, x, position_ids):
        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
        position_ids_expanded = position_ids[:, None, :].float()
        freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
        emb = torch.cat((freqs, freqs), dim=-1)
        emb = emb.unsqueeze(1)
        cos = emb.cos()
        sin = emb.sin()
        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)


def rotate_half(x):
    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):
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    return q_embed, k_embed


class NemotronDenseMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.up_proj = nn.Linear(config.hidden_size, config.intermediate_size, bias=False)
        self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=False)
        self.act_fn = ACT2FN[config.hidden_act]

    def forward(self, x):
        return self.down_proj(self.act_fn(self.up_proj(x)))


class NemotronDenseAttention(nn.Module):
    def __init__(self, config, layer_idx=None):
        super().__init__()
        self.layer_idx = layer_idx
        self.hidden_size = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = getattr(config, "head_dim", None) or self.hidden_size // self.num_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.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)

        self.rotary_emb = NemotronDenseRotaryEmbedding(
            self.head_dim,
            max_position_embeddings=config.max_position_embeddings,
            base=config.rope_parameters["rope_theta"],
        )

    def forward(self, hidden_states, attention_mask=None, position_ids=None, past_key_values=None, use_cache=False, **kwargs):
        bsz, q_len, _ = hidden_states.size()

        query_states = self.q_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
        key_states = self.k_proj(hidden_states).view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
        value_states = self.v_proj(hidden_states).view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)

        past_len = past_key_values.get_seq_length(self.layer_idx) if past_key_values is not None else 0
        if position_ids is None:
            position_ids = torch.arange(past_len, past_len + q_len, dtype=torch.long, device=hidden_states.device).unsqueeze(0)

        cos, sin = self.rotary_emb(value_states, position_ids)
        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

        if past_key_values is not None:
            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)

        kv_seq_len = key_states.shape[-2]
        key_states = repeat_kv(key_states, self.num_key_value_groups)
        value_states = repeat_kv(value_states, self.num_key_value_groups)

        if attention_mask is not None:
            if attention_mask.dim() == 2:
                attention_mask = attention_mask[:, None, None, :kv_seq_len].to(torch.bool)
                if q_len > 1:
                    causal = torch.tril(
                        torch.ones(q_len, kv_seq_len, dtype=torch.bool, device=hidden_states.device),
                        diagonal=kv_seq_len - q_len,
                    )
                    attention_mask = attention_mask & causal[None, None, :, :]
            else:
                attention_mask = attention_mask[:, :, :, :kv_seq_len]

        is_causal = attention_mask is None and q_len > 1

        attn_output = torch.nn.functional.scaled_dot_product_attention(
            query_states, key_states, value_states,
            attn_mask=attention_mask, dropout_p=0.0, is_causal=is_causal,
        )

        attn_output = attn_output.transpose(1, 2).contiguous()
        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
        attn_output = self.o_proj(attn_output)
        return attn_output, None


class NemotronDenseDecoderLayer(nn.Module):
    def __init__(self, config, layer_idx=None):
        super().__init__()
        self.self_attn = NemotronDenseAttention(config, layer_idx=layer_idx)
        self.mlp = NemotronDenseMLP(config)
        self.input_layernorm = NemotronDenseRMSNorm(config.hidden_size, eps=config.norm_eps)
        self.post_attention_layernorm = NemotronDenseRMSNorm(config.hidden_size, eps=config.norm_eps)

    def forward(self, hidden_states, attention_mask=None, position_ids=None, past_key_values=None, use_cache=False, **kwargs):
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)
        hidden_states, _ = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            use_cache=use_cache,
        )
        hidden_states = residual + hidden_states

        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 NemotronDenseModel(PreTrainedModel):
    config_class = NemotronDenseConfig
    base_model_prefix = "model"

    def __init__(self, config: NemotronDenseConfig):
        super().__init__(config)
        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
        self.layers = nn.ModuleList(
            [NemotronDenseDecoderLayer(config, layer_idx=i) for i in range(config.num_hidden_layers)]
        )
        self.norm = NemotronDenseRMSNorm(config.hidden_size, eps=config.norm_eps)
        self.post_init()

    def _init_weights(self, module):
        if isinstance(module, NemotronDenseRotaryEmbedding):
            inv_freq = 1.0 / (module.base ** (torch.arange(0, module.dim, 2, dtype=torch.int64).float() / module.dim))
            try:
                import transformers.initialization as init
                init.copy_(module.inv_freq, inv_freq)
            except (ImportError, AttributeError):
                module.inv_freq.copy_(inv_freq)

    def get_input_embeddings(self):
        return self.embed_tokens

    def set_input_embeddings(self, new_embeddings):
        self.embed_tokens = new_embeddings

    def forward(self, input_ids=None, attention_mask=None, position_ids=None, past_key_values=None, use_cache=None, return_dict=None, **kwargs):
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        if use_cache and past_key_values is None:
            past_key_values = DynamicCache()

        hidden_states = self.embed_tokens(input_ids)

        for decoder_layer in self.layers:
            hidden_states = decoder_layer(
                hidden_states,
                attention_mask=attention_mask,
                position_ids=position_ids,
                past_key_values=past_key_values,
                use_cache=use_cache,
            )

        hidden_states = self.norm(hidden_states)

        if not return_dict:
            return tuple(v for v in [hidden_states, past_key_values] if v is not None)

        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values,
        )


class NemotronDenseForCausalLM(PreTrainedModel, GenerationMixin):
    config_class = NemotronDenseConfig
    base_model_prefix = "model"

    def __init__(self, config: NemotronDenseConfig):
        super().__init__(config)
        self.model = NemotronDenseModel(config)
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
        self.post_init()

    def forward(self, input_ids=None, attention_mask=None, position_ids=None, past_key_values=None, labels=None, use_cache=None, return_dict=None, **kwargs):
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        outputs = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            use_cache=use_cache,
            return_dict=return_dict,
            **kwargs,
        )

        hidden_states = outputs[0]
        logits = self.lm_head(hidden_states)

        loss = None
        if labels is not None:
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            loss_fct = nn.CrossEntropyLoss()
            loss = loss_fct(shift_logits.view(-1, self.config.vocab_size), shift_labels.view(-1))

        if not return_dict:
            output = (logits,) + outputs[1:]
            return ((loss,) + output) if loss is not None else output

        return CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
        )

    def get_input_embeddings(self):
        return self.model.get_input_embeddings()

    def set_input_embeddings(self, new_embeddings):
        return self.model.set_input_embeddings(new_embeddings)

    def get_output_embeddings(self):
        return self.lm_head

    def set_output_embeddings(self, new_embeddings):
        self.lm_head = new_embeddings

    def get_decoder(self):
        return self.model

    def set_decoder(self, decoder):
        self.model = decoder

    # `prepare_inputs_for_generation` intentionally not overridden:
    # transformers.GenerationMixin's default already handles next_sequence_length,
    # inputs_embeds first-iteration injection, left-padded position_ids, compilable
    # caches, and past_key_values forwarding. Inheriting it gets us all of those
    # correctly without us having to keep our override in sync with HF.