# coding=utf-8 # Copyright 2025 NVIDIA Corporation. All rights reserved. """ PyTorch Nemotron-Flash model.""" import inspect import math import copy import warnings from typing import Any, Dict, List, Optional, Tuple, Union import time import numpy as np import os import torch import torch.nn.functional as F import torch.utils.checkpoint from torch import nn from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss torch._inductor.config.max_autotune_gemm_backends = ["aten"] from transformers.activations import ACT2FN from transformers.cache_utils import Cache, DynamicCache from transformers.modeling_outputs import ( MoeCausalLMOutputWithPast, MoeModelOutputWithPast, ) from transformers.modeling_utils import PreTrainedModel from transformers.generation import GenerationMixin try: from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update except ImportError: pass from transformers.utils import ( is_flash_attn_greater_or_equal_2_10, logging, replace_return_docstrings, ) from .configuration_nemotron_flash import NemotronFlashConfig import math from flash_attn import flash_attn_func, flash_attn_varlen_func from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input # noqa _flash_supports_window_size = "window_size" in list(inspect.signature(flash_attn_func).parameters) from einops import rearrange, repeat, reduce, pack, unpack from .fused_mha_with_cache import fused_mha_interface from .mamba2 import Mamba2 from mamba_ssm.utils.generation import InferenceParams from .delta_net import Cache as fla_cache from .delta_net import DeltaNet import torch._dynamo torch._dynamo.config.suppress_errors = True from torch.cuda import CUDAGraph logger = logging.get_logger(__name__) _CONFIG_FOR_DOC = "NemotronFlashConfig" class NemotronFlashRMSNorm(nn.Module): def __init__(self, hidden_size, learnable_weight=True, eps=1e-6): super().__init__() if learnable_weight: self.weight = nn.Parameter(torch.ones(hidden_size)) else: self.weight = None 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) if self.weight is not None: return self.weight * hidden_states.to(input_dtype) else: return hidden_states.to(input_dtype) class LlamaRotaryEmbedding(nn.Module): def __init__(self, config, dim, base=10000, device=None, scaling_factor=1.0): super().__init__() self.scaling_factor = scaling_factor self.dim = dim self.base = base self.config = config self.rope_type = config.rope_type self.factor = 2 max_position_embeddings = self.config.max_position_embeddings if config.rope_type is None or config.rope_type == "default": inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float().to(device) / self.dim)) self.max_seq_len_cached = max_position_embeddings elif config.rope_type == 'ntk': assert self.config.orig_max_position_embeddings is not None orig_max_position_embeddings = self.config.orig_max_position_embeddings base = base * ((self.factor * max_position_embeddings / orig_max_position_embeddings) - (self.factor - 1)) ** (self.dim / (self.dim - 2)) inv_freq = 1.0 / (base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float().to(device) / self.dim)) self.max_seq_len_cached = orig_max_position_embeddings elif config.rope_type == 'dynamic_ntk': inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float().to(device) / self.dim)) self.original_inv_freq = inv_freq self.max_seq_len_cached = self.config.orig_max_position_embeddings else: raise ValueError(f"Not support rope_type: {config.rope_type}") self.register_buffer("inv_freq", inv_freq, persistent=False) def _dynamic_frequency_update(self, position_ids, device): """ dynamic RoPE layers should recompute `inv_freq` in the following situations: 1 - growing beyond the cached sequence length (allow scaling) 2 - the current sequence length is in the original scale (avoid losing precision with small sequences) """ seq_len = torch.max(position_ids) + 1 if seq_len > self.max_seq_len_cached: # growth base = self.base * ((self.factor * seq_len / self.config.orig_max_position_embeddings) - (self.factor - 1)) ** (self.dim / (self.dim - 2)) inv_freq = 1.0 / (base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float().to(device) / self.dim)) self.register_buffer("inv_freq", inv_freq, persistent=False) self.max_seq_len_cached = seq_len if seq_len < self.config.orig_max_position_embeddings and self.max_seq_len_cached > self.config.orig_max_position_embeddings: # reset self.register_buffer("inv_freq", self.original_inv_freq, persistent=False) self.max_seq_len_cached = self.config.orig_max_position_embeddings @torch.no_grad() def forward(self, x, position_ids): if self.rope_type == 'dynamic_ntk': self._dynamic_frequency_update(position_ids, device=x.device) inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1) position_ids_expanded = position_ids[:, None, :].float() device_type = x.device.type device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu" with torch.autocast(device_type=device_type, enabled=False): freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2) emb = torch.cat((freqs, freqs), dim=-1) cos = emb.cos() sin = emb.sin() return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype) 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.""" cos = cos.unsqueeze(unsqueeze_dim) sin = sin.unsqueeze(unsqueeze_dim) if q is not None: q_embed = (q * cos) + (rotate_half(q) * sin) else: q_embed = None if k is not None: k_embed = (k * cos) + (rotate_half(k) * sin) else: k_embed = None return q_embed, k_embed 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 AttentionDynamicCache(DynamicCache): def __init__(self, config, batch_size, dtype=torch.float16, device=None, layer_type=None): self.dtype = dtype self.key_cache = [torch.tensor([[]] * batch_size, device=device) for _ in range(config.num_hidden_layers)] self.value_cache = [torch.tensor([[]] * batch_size, device=device) for _ in range(config.num_hidden_layers)] def update( self, key_states: torch.Tensor, value_states: torch.Tensor, layer_idx: int, cache_kwargs: Optional[Dict[str, Any]] = None, ) -> Tuple[torch.Tensor, torch.Tensor]: if self.key_cache[layer_idx].shape[-1] == 0: self.key_cache[layer_idx] = key_states self.value_cache[layer_idx] = value_states else: self.key_cache[layer_idx] = torch.cat([self.key_cache[layer_idx], key_states], dim=2) self.value_cache[layer_idx] = torch.cat([self.value_cache[layer_idx], value_states], dim=2) return self.key_cache[layer_idx], self.value_cache[layer_idx] def get_seq_length(self, layer_idx=None) -> int: if layer_idx is None: max_key_len = max(cache.shape[-2] for cache in self.key_cache) return max_key_len if self.key_cache[layer_idx].shape[-1] == 0: return 0 return self.key_cache[layer_idx].shape[-2] # Adapted from transformers.models.mistral.modeling_mistral.MistralAttention class NemotronFlashAttention(nn.Module): def __init__(self, config: NemotronFlashConfig, layer_idx: Optional[int] = None, input_hidden_size=None, output_hidden_size=None): super().__init__() self.config = config self.layer_idx = layer_idx if layer_idx is None: logger.warning_once( f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will " "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` " "when creating this class." ) self.hidden_size = config.attn_hidden_size if config.attn_hidden_size > 0 else config.hidden_size self.num_heads = config.num_attention_heads self.head_dim = self.hidden_size // self.num_heads self.max_position_embeddings = config.max_position_embeddings self.rope_theta = config.rope_theta self.kq_head_dim = config.kq_head_dim if config.kq_head_dim > 0 else self.head_dim self.v_head_dim = config.v_head_dim if config.v_head_dim > 0 else self.head_dim self.num_key_value_heads = config.num_key_value_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.is_causal = True self.attention_dropout = config.attention_dropout if (self.head_dim * self.num_heads) != self.hidden_size and self.kq_head_dim == self.head_dim: raise ValueError( f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}" f" and `num_heads`: {self.num_heads})." ) self.q_proj = nn.Linear(self.hidden_size if input_hidden_size is None else input_hidden_size, self.num_heads * self.kq_head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size if input_hidden_size is None else input_hidden_size, self.num_key_value_heads * self.kq_head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size if input_hidden_size is None else input_hidden_size, self.num_key_value_heads * self.v_head_dim, bias=False) if output_hidden_size is None: output_hidden_size = self.hidden_size self.o_proj = nn.Linear(self.num_heads * self.v_head_dim, output_hidden_size, bias=False) if self.config.kq_norm == "rms": self.k_norm = NemotronFlashRMSNorm(self.kq_head_dim) self.q_norm = NemotronFlashRMSNorm(self.kq_head_dim) elif self.config.kq_norm == "none": self.k_norm = None self.q_norm = None else: raise NotImplementedError(f"Unknown kq_norm: {self.config.kq_norm}") if self.config.rope: self._init_rope() def _init_rope(self): self.rotary_emb = LlamaRotaryEmbedding( config=self.config, dim=self.kq_head_dim, base=self.rope_theta, device=torch.device("cuda"), ) def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int): return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous() 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, output_attentions: bool = False, use_cache: bool = False, use_swa=False, query_states = None, key_states=None, value_states=None, **kwargs, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: raise NotImplementedError("NemotronFlashAttention is an abstract class. Use one of the subclasses.") def _get_unpad_data(attention_mask): seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32) indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten() max_seqlen_in_batch = seqlens_in_batch.max().item() cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0)) return ( indices, cu_seqlens, max_seqlen_in_batch, ) # Adapted from transformers.models.mistral.modeling_mistral.MistralFlashAttention2 class NemotronFlashFlashAttention2(NemotronFlashAttention): def __init__(self, *args, **kwargs): super().__init__(*args, **kwargs) self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10() def forward( self, hidden_states: torch.Tensor = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, use_swa=False, query_states = None, key_states=None, value_states=None, **kwargs, ): if "padding_mask" in kwargs: warnings.warn( "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`" ) # overwrite attention_mask with padding_mask attention_mask = kwargs.pop("padding_mask") bsz, q_len, _ = hidden_states.size() query_states = self.q_proj(hidden_states) query_states = query_states.view(bsz, q_len, self.num_heads, self.kq_head_dim).transpose(1, 2).contiguous() if self.q_norm is not None: query_states = self.q_norm(query_states) if self.config.rope: cos, sin = self.rotary_emb(hidden_states, position_ids) query_states, _ = apply_rotary_pos_emb(query_states, None, cos, sin) key_states = self.k_proj(hidden_states) value_states = self.v_proj(hidden_states) key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.kq_head_dim).transpose(1, 2) value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.v_head_dim).transpose(1, 2) if self.k_norm is not None: key_states = self.k_norm(key_states) if self.config.rope: _, key_states = apply_rotary_pos_emb(None, key_states, cos, sin) kv_seq_len = key_states.shape[-2] if past_key_value is not None: if self.layer_idx is None: raise ValueError( f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} " "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class " "with a layer index." ) kv_seq_len += past_key_value.get_seq_length(self.layer_idx) use_sliding_windows = ( _flash_supports_window_size and getattr(self.config, "sliding_window", None) is not None and kv_seq_len > self.config.sliding_window and use_swa ) if not _flash_supports_window_size: logger.warning_once( "The current flash attention version does not support sliding window attention, for a more memory efficient implementation" " make sure to upgrade flash-attn library." ) swa_processed_flag = False if past_key_value is not None and use_cache: kv_layer_idx = self.layer_idx cache_has_contents = past_key_value.get_seq_length(kv_layer_idx) > 0 if ( getattr(self.config, "sliding_window", None) is not None and kv_seq_len > self.config.sliding_window and cache_has_contents and use_swa ): slicing_tokens = 1 - self.config.sliding_window past_key = past_key_value[kv_layer_idx][0] past_value = past_key_value[kv_layer_idx][1] past_key = past_key[:, :, slicing_tokens:, :].contiguous() past_value = past_value[:, :, slicing_tokens:, :].contiguous() past_key_value.key_cache[kv_layer_idx] = past_key past_key_value.value_cache[kv_layer_idx] = past_value if attention_mask is not None: attention_mask = attention_mask[:, slicing_tokens:] attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1) swa_processed_flag = True key_states, value_states = past_key_value.update(key_states, value_states, kv_layer_idx) key_states_no_repeat = key_states value_states_no_repeat = value_states key_states = repeat_kv(key_states, self.num_key_value_groups) value_states = repeat_kv(value_states, self.num_key_value_groups) dropout_rate = 0.0 if not self.training else self.attention_dropout input_dtype = query_states.dtype if input_dtype == torch.float32: if torch.is_autocast_enabled(): target_dtype = torch.get_autocast_gpu_dtype() elif hasattr(self.config, "_pre_quantization_dtype"): target_dtype = self.config._pre_quantization_dtype else: target_dtype = self.q_proj.weight.dtype logger.warning_once( f"The input hidden states seems to be silently casted in float32, this might be related to" f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in" f" {target_dtype}." ) query_states = query_states.to(target_dtype) key_states = key_states.to(target_dtype) value_states = value_states.to(target_dtype) # Reashape to the expected shape for Flash Attention query_states = query_states.transpose(1, 2) # (batch, slen, num_heads, head_dim) key_states = key_states.transpose(1, 2) # (batch, slen, num_heads, head_dim) value_states = value_states.transpose(1, 2) # (batch, slen, num_heads, head_dim) attn_output = self._flash_attention_forward( query_states, key_states, value_states, attention_mask, q_len, dropout=dropout_rate, use_sliding_windows=use_sliding_windows and not swa_processed_flag, ) v_dim = value_states.shape[-2] * value_states.shape[-1] attn_output = attn_output.reshape(-1, q_len, v_dim).contiguous() attn_output = self.o_proj(attn_output) if not output_attentions: attn_weights = None return attn_output, attn_weights, past_key_value, (key_states_no_repeat, value_states_no_repeat) def _flash_attention_forward( self, query_states, key_states, value_states, attention_mask, query_length, dropout=0.0, softmax_scale=None, use_sliding_windows=False, ): """ Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token first unpad the input, then computes the attention scores and pad the final attention scores. Args: query_states (`torch.Tensor`): Input query states to be passed to Flash Attention API key_states (`torch.Tensor`): Input key states to be passed to Flash Attention API value_states (`torch.Tensor`): Input value states to be passed to Flash Attention API attention_mask (`torch.Tensor`): The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the position of padding tokens and 1 for the position of non-padding tokens. dropout (`int`, *optional*): Attention dropout softmax_scale (`float`, *optional*): The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim) use_sliding_windows (`bool`, *optional*): Whether to activate sliding window attention. """ if not self._flash_attn_uses_top_left_mask: causal = self.is_causal else: # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__. causal = self.is_causal and query_length != 1 if attention_mask is not None: batch_size = query_states.shape[0] query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input( query_states, key_states, value_states, attention_mask, query_length ) cu_seqlens_q, cu_seqlens_k = cu_seq_lens max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens if not use_sliding_windows: attn_output_unpad = flash_attn_varlen_func( query_states, key_states, value_states, cu_seqlens_q=cu_seqlens_q, cu_seqlens_k=cu_seqlens_k, max_seqlen_q=max_seqlen_in_batch_q, max_seqlen_k=max_seqlen_in_batch_k, dropout_p=dropout, softmax_scale=softmax_scale, causal=causal, ) else: attn_output_unpad = flash_attn_varlen_func( query_states, key_states, value_states, cu_seqlens_q=cu_seqlens_q, cu_seqlens_k=cu_seqlens_k, max_seqlen_q=max_seqlen_in_batch_q, max_seqlen_k=max_seqlen_in_batch_k, dropout_p=dropout, softmax_scale=softmax_scale, causal=causal, window_size=(self.config.sliding_window, self.config.sliding_window), ) attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length) else: if not use_sliding_windows: attn_output = flash_attn_func( query_states, key_states, value_states, dropout, softmax_scale=softmax_scale, causal=causal, ) else: attn_output = flash_attn_func( query_states, key_states, value_states, dropout, softmax_scale=softmax_scale, causal=causal, window_size=(self.config.sliding_window, self.config.sliding_window), ) return attn_output def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length): batch_size, kv_seq_len, num_heads, head_dim = key_layer.shape # On the first iteration we need to properly re-create the padding mask # by slicing it on the proper place if kv_seq_len != attention_mask.shape[-1]: attention_mask_num_tokens = attention_mask.shape[-1] attention_mask = attention_mask[:, attention_mask_num_tokens - kv_seq_len :] indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask) if not self.training and not type(key_layer) == torch.Tensor: ## this is for handling Mamba2 with output type key_layer = torch.tensor(key_layer.clone()) value_layer = torch.tensor(value_layer.clone()) query_layer = torch.tensor(query_layer.clone()) key_layer = index_first_axis(key_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k) value_layer = index_first_axis(value_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k) if query_length == kv_seq_len: query_layer = index_first_axis( query_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k ) cu_seqlens_q = cu_seqlens_k max_seqlen_in_batch_q = max_seqlen_in_batch_k indices_q = indices_k elif query_length == 1: max_seqlen_in_batch_q = 1 cu_seqlens_q = torch.arange( batch_size + 1, dtype=torch.int32, device=query_layer.device ) # There is a memcpy here, that is very bad. indices_q = cu_seqlens_q[:-1] query_layer = query_layer.squeeze(1) else: # The -q_len: slice assumes left padding. attention_mask = attention_mask[:, -query_length:] query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask) return ( query_layer, key_layer, value_layer, indices_q, (cu_seqlens_q, cu_seqlens_k), (max_seqlen_in_batch_q, max_seqlen_in_batch_k), ) class NemotronFlashSDPAAttention(nn.Module): def __init__(self, config, layer_idx: int, reuse_kv=False): 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 = self.config.sliding_window if self.layer_idx not in self.config.global_attn_idx else None self.rotary_emb = NemotronFlashRotaryEmbedding(config=config) def forward( self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor], position_ids: Optional[torch.LongTensor] = None, past_key_value: Optional[Cache] = None, **kwargs, ) -> tuple[torch.Tensor, Optional[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 = self.rotary_emb(hidden_states, position_ids) query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin) if past_key_value is not None: past_seen_tokens = past_key_value.get_seq_length() cache_position = torch.arange( past_seen_tokens, past_seen_tokens + hidden_states.shape[1], device=hidden_states.device ) 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) attention_interface = ALL_ATTENTION_FUNCTIONS['flash_attention_2'] attn_output, attn_weights = attention_interface( self, query_states, key_states, value_states, attention_mask, dropout=0.0 if not self.training else self.attention_dropout, scaling=self.scaling, sliding_window=self.sliding_window, **kwargs, ) attn_output = attn_output.reshape(*input_shape, -1).contiguous() attn_output = self.o_proj(attn_output) return attn_output, attn_weights, past_key_value, (key_states, value_states) class NemotronFlashRotaryEmbedding(nn.Module): def __init__(self, config, device=None): super().__init__() if hasattr(config, "rope_scaling") and config.rope_scaling is not None: 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 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) ## Interface to use TRTLLM AutoDeploy attention kernel, which enables CUDA Graph capture class NemotronFlashFusedMHA(NemotronFlashAttention): def __init__(self, *args, **kwargs): super().__init__(*args, **kwargs) self.fused_mha_interface = fused_mha_interface def init_kv_cache(self, max_batch_size, max_seq_len, page_size=-1): if hasattr(self, 'k_cache'): del self.k_cache del self.v_cache if hasattr(self, 'page_table') and self.page_table is not None: del self.page_table import gc gc.collect() torch.cuda.empty_cache() if page_size is not None and page_size > 0: batch_max_pages = (max_seq_len + page_size - 1) // page_size cache_max_pages = (max_batch_size * max_seq_len + page_size - 1) // page_size self.k_cache = torch.zeros(cache_max_pages, page_size, self.num_key_value_heads, self.kq_head_dim).to(self.q_proj.weight) self.v_cache = torch.zeros(cache_max_pages, page_size, self.num_key_value_heads, self.v_head_dim).to(self.q_proj.weight) self.page_table = torch.zeros(max_batch_size, batch_max_pages, device=self.q_proj.weight.device, dtype=torch.int32) else: self.k_cache = torch.zeros(max_batch_size, max_seq_len, self.num_key_value_heads, self.kq_head_dim).to(self.q_proj.weight) self.v_cache = torch.zeros(max_batch_size, max_seq_len, self.num_key_value_heads, self.v_head_dim).to(self.q_proj.weight) self.page_table = None self.max_seq_len = max_seq_len def reset_kv_cache(self): self.k_cache = self.k_cache.zero_() self.v_cache = self.v_cache.zero_() if self.page_table is not None: self.page_table = self.page_table.zero_() def forward( self, hidden_states: torch.Tensor = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, use_swa=False, query_states = None, key_states=None, value_states=None, **kwargs, ): if not hasattr(self, 'k_cache'): self.init_kv_cache(max_batch_size=1, max_seq_len=8000) bsz, q_len, _ = hidden_states.size() query_states = self.q_proj(hidden_states) query_states = query_states.view(bsz, q_len, self.num_heads, self.kq_head_dim).transpose(1, 2).contiguous() if self.q_norm is not None: query_states = self.q_norm(query_states) if self.config.rope: cos, sin = self.rotary_emb(hidden_states, position_ids) query_states, _ = apply_rotary_pos_emb(query_states, None, cos, sin) key_states = self.k_proj(hidden_states) value_states = self.v_proj(hidden_states) key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.kq_head_dim).transpose(1, 2) value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.v_head_dim).transpose(1, 2) if self.k_norm is not None: key_states = self.k_norm(key_states) if self.config.rope: _, key_states = apply_rotary_pos_emb(None, key_states, cos, sin) key_states_no_repeat = key_states value_states_no_repeat = value_states query_states = query_states.transpose(1, 2) # (batch, slen, num_heads, head_dim) key_states = key_states.transpose(1, 2) # (batch, slen, num_kv_heads, head_dim) value_states = value_states.transpose(1, 2) # (batch, slen, num_kv_heads, head_dim) if self.k_cache.device != query_states.device: self.k_cache = self.k_cache.to(query_states) self.v_cache = self.v_cache.to(query_states) attn_output = self.fused_mha_interface( query_states, key_states, value_states, k_cache=self.k_cache, v_cache=self.v_cache, page_table=self.page_table, max_seq_len=self.max_seq_len, position_ids=position_ids, ) v_dim = query_states.shape[-2] * value_states.shape[-1] attn_output = attn_output.reshape(bsz, q_len, v_dim).contiguous() attn_output = self.o_proj(attn_output) if not output_attentions: attn_weights = None return attn_output, attn_weights, past_key_value, (key_states_no_repeat, value_states_no_repeat) JAMBA_ATTENTION_CLASSES = { "flash_attention_2": NemotronFlashFlashAttention2, "fused_mha": NemotronFlashFusedMHA, "sdpa": NemotronFlashSDPAAttention, } class NemotronFlashMLP(nn.Module): def __init__(self, config: NemotronFlashConfig, layer_idx: int): super().__init__() self.config = config self.act_fn_name = config.mlp_hidden_act self.act_fn = ACT2FN[self.act_fn_name] if config.ffn_expand_ratio is not None: self.ffn_dim = int(config.ffn_expand_ratio * config.hidden_size) // 128 * 128 else: self.ffn_dim = config.intermediate_size self.hidden_dim = config.hidden_size self.layer_idx = layer_idx if self.act_fn_name == "silu": self.gate_proj = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False) self.down_proj = nn.Linear(self.ffn_dim, self.hidden_dim, bias=False) self.up_proj = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False) def forward(self, x): if self.act_fn_name == "silu": output = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x)) elif self.act_fn_name == "relu2": output = self.down_proj(self.act_fn(self.up_proj(x))) else: raise NotImplementedError(f"No such hidden_act: {self.act_fn_name}") return output class NemotronFlashAttentionDecoderLayer(nn.Module): def __init__(self, config: NemotronFlashConfig, layer_idx: int,): super().__init__() self.config = config self.layer_idx = layer_idx self.self_attn = JAMBA_ATTENTION_CLASSES[config.attn_implementation](config, layer_idx) if self.config.intermediate_size > 0: self.ffn = NemotronFlashMLP(config, layer_idx=layer_idx) self.pre_ffn_layernorm = NemotronFlashRMSNorm(config.hidden_size, eps=config.rms_norm_eps) else: self.ffn = None self.pre_ffn_layernorm = None self.input_layernorm = NemotronFlashRMSNorm(config.hidden_size, eps=config.rms_norm_eps) def forward( self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_value: Optional[Tuple[torch.Tensor]] = None, output_attentions: Optional[bool] = False, use_cache: Optional[bool] = False, use_swa=False, **kwargs, ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]: if "padding_mask" in kwargs: warnings.warn( "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`" ) if position_ids is not None and position_ids.shape[1] != hidden_states.shape[1]: position_ids = torch.arange(hidden_states.shape[1], device=hidden_states.device).unsqueeze(0) residual = hidden_states if self.input_layernorm is not None: hidden_states = self.input_layernorm(hidden_states) hidden_states, self_attn_weights, present_key_value, current_kv = self.self_attn( hidden_states=hidden_states, attention_mask=attention_mask, position_ids=position_ids, past_key_value=past_key_value, output_attentions=output_attentions, use_cache=use_cache, use_swa=use_swa, ) hidden_states = residual + hidden_states if self.ffn is not None: residual = hidden_states if self.pre_ffn_layernorm is not None: hidden_states = self.pre_ffn_layernorm(hidden_states) hidden_states = self.ffn(hidden_states) hidden_states = residual + hidden_states outputs = (hidden_states,) if output_attentions: outputs += (self_attn_weights,) if use_cache: outputs += (present_key_value,) outputs += (current_kv,) return outputs class FFNDecoderLayer(nn.Module): def __init__(self, config: NemotronFlashConfig, layer_idx: int): super().__init__() self.config = config self.layer_idx = layer_idx self.ffn = NemotronFlashMLP(config, layer_idx=layer_idx) self.pre_ffn_layernorm = NemotronFlashRMSNorm(config.hidden_size, eps=config.rms_norm_eps) def forward( self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_value: Optional[Tuple[torch.Tensor]] = None, output_attentions: Optional[bool] = False, use_cache: Optional[bool] = False, use_swa=False, **kwargs, ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]: if "padding_mask" in kwargs: warnings.warn( "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`" ) residual = hidden_states if self.pre_ffn_layernorm is not None: hidden_states = self.pre_ffn_layernorm(hidden_states) hidden_states = self.ffn(hidden_states) hidden_states = residual + hidden_states outputs = (hidden_states,) if output_attentions: outputs += (None,) if use_cache: outputs += (None,) return outputs class NemotronFlashMambaDecoderLayer(nn.Module): def __init__(self, config: NemotronFlashConfig, layer_idx: int): super().__init__() self.config = config self.layer_idx = layer_idx self.mamba = Mamba2(config=config, layer_idx=layer_idx) self.intermediate_size = config.intermediate_size if self.intermediate_size > 0: self.ffn = NemotronFlashMLP(config, layer_idx=layer_idx) self.pre_ffn_layernorm = NemotronFlashRMSNorm(config.hidden_size, eps=config.rms_norm_eps) else: self.ffn = None self.pre_ffn_layernorm = None self.input_layernorm = NemotronFlashRMSNorm(config.hidden_size, eps=config.rms_norm_eps) def forward( self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_value: Optional[AttentionDynamicCache] = None, output_attentions: Optional[bool] = False, use_cache: Optional[bool] = False, use_swa=False, mamba_inference_params=None, **kwargs, ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]: if "padding_mask" in kwargs: warnings.warn( "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`" ) if position_ids is not None and position_ids.shape[1] != hidden_states.shape[1]: position_ids = torch.arange(hidden_states.shape[1], device=hidden_states.device).unsqueeze(0) residual = hidden_states if self.input_layernorm is not None: hidden_states = self.input_layernorm(hidden_states) hidden_states, present_key_value = self.mamba( hidden_states=hidden_states, past_key_value=past_key_value, attention_mask=attention_mask, inference_params=mamba_inference_params, ) attn_key_value = None hidden_states = residual + hidden_states if self.intermediate_size > 0: residual = hidden_states if self.pre_ffn_layernorm is not None: hidden_states = self.pre_ffn_layernorm(hidden_states) hidden_states = self.ffn(hidden_states) hidden_states = residual + hidden_states outputs = (hidden_states,) if use_cache: outputs += (present_key_value,) outputs += (attn_key_value,) return outputs def _get_past_seqlen(self, past_key_value, seqlen): if past_key_value is None: return seqlen past_seqlen = past_key_value.get_seq_length(self.layer_idx) if past_seqlen == 0: return seqlen return past_seqlen class NemotronFlashHybridDecoderLayer(nn.Module): def __init__(self, config: NemotronFlashConfig, layer_idx: int): super().__init__() self.config = config self.layer_idx = layer_idx if config.hybrid_decoder_layer == 'mamba': self.mamba = Mamba2(config=config, layer_idx=layer_idx) if config.hybrid_decoder_layer == 'deltanet': if config.layer_types is not None: deltanet_idx = sum(1 for i in range(layer_idx) if config.layer_types[i] == 'deltanet') else: deltanet_idx = layer_idx self.gla = DeltaNet(hidden_size=config.hidden_size, num_heads=config.num_attention_heads, layer_idx=deltanet_idx, config=self.config) else: raise ValueError(f"Not supported: {config.hybrid_decoder_layer}") self.config = config if self.config.intermediate_size > 0: self.ffn = NemotronFlashMLP(config, layer_idx=layer_idx) self.pre_ffn_layernorm = NemotronFlashRMSNorm(config.hidden_size, eps=config.rms_norm_eps) else: self.ffn = None self.pre_ffn_layernorm = None self.input_layernorm = NemotronFlashRMSNorm(config.hidden_size, eps=config.rms_norm_eps) def forward( self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_value: Optional[AttentionDynamicCache] = None, output_attentions: Optional[bool] = False, use_cache: Optional[bool] = False, fla_past_key_values = None, mamba_inference_params = None, use_swa=False, **kwargs, ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]: if "padding_mask" in kwargs: warnings.warn( "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`" ) residual = hidden_states hidden_states = self.input_layernorm(hidden_states) if self.config.hybrid_decoder_layer == 'mamba': hybrid_op_hidden_states, mamba_present_key_value = self.mamba( hidden_states=hidden_states, past_key_value=past_key_value, attention_mask=attention_mask, inference_params=mamba_inference_params, ) else: hybrid_op_hidden_states, _, fla_past_key_values = self.gla( hidden_states=hidden_states, attention_mask=attention_mask, past_key_values=fla_past_key_values, use_cache=use_cache, ) self_attn_weights = self_attn_present_key_value = current_kv = None hidden_states = residual + hybrid_op_hidden_states if self.ffn is not None: residual = hidden_states hidden_states = self.pre_ffn_layernorm(hidden_states) hidden_states = self.ffn(hidden_states) hidden_states = residual + hidden_states outputs = (hidden_states,) if output_attentions: outputs += (self_attn_weights,) if use_cache: outputs += (self_attn_present_key_value,) outputs += (current_kv,) return outputs # Adapted from transformers.models.mistral.modeling_mistral.MistralPreTrainedModel class NemotronFlashPreTrainedModel(PreTrainedModel): config_class = NemotronFlashConfig base_model_prefix = "model" supports_gradient_checkpointing = True _no_split_modules = ["NemotronFlashAttentionDecoderLayer", "NemotronFlashMambaDecoderLayer"] _skip_keys_device_placement = "past_key_values" _supports_flash_attn_2 = True _supports_sdpa = True _supports_cache_class = True def _init_weights(self, module): std = self.config.initializer_range if isinstance(module, (nn.Linear, nn.Conv1d)): 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_() # Adapted from transformers.models.mistral.modeling_mistral.MistralModel class NemotronFlashModel(NemotronFlashPreTrainedModel): def __init__(self, config: NemotronFlashConfig): super().__init__(config) config.attn_implementation = config.attn_implementation_new config._attn_implementation = config.attn_implementation_new self.config = config self.padding_idx = config.pad_token_id self.vocab_size = config.vocab_size self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx) decoder_layers = [] layer_type = [] for i in range(config.num_hidden_layers): if config.layer_types[i] in ['deltanet']: layer_type.append('m') config_new = copy.deepcopy(config) config_new.hybrid_decoder_layer = 'deltanet' decoder_layer = NemotronFlashHybridDecoderLayer(config_new, layer_idx=i) elif config.layer_types[i] in ['m', 'm2']: layer_type.append('m') decoder_layer = NemotronFlashMambaDecoderLayer(config, layer_idx=i) elif config.layer_types[i] == 'a': layer_type.append('a') decoder_layer = NemotronFlashAttentionDecoderLayer(config, layer_idx=i) elif config.layer_types[i] == 'f': layer_type.append('a') decoder_layer = FFNDecoderLayer(config, layer_idx=i) else: raise ValueError(f"Unsupported layer type {config.layer_types[i]}") decoder_layers.append(decoder_layer) config.layer_type = layer_type if config.sliding_window is not None: self.sliding_window = config.sliding_window self.global_attn_idx = config.global_attn_idx else: self.sliding_window = None self.global_attn_idx = None self.layers = nn.ModuleList(decoder_layers) self._attn_implementation = config.attn_implementation self.final_layernorm = NemotronFlashRMSNorm(config.hidden_size, eps=config.rms_norm_eps) if self.config.num_memory_tokens > 0: self.memory_tokens = nn.Parameter(torch.randn(self.config.num_memory_tokens, self.config.hidden_size)) self.gradient_checkpointing = False self.post_init() self.has_previous_state = False def get_input_embeddings(self): return self.embed_tokens def set_input_embeddings(self, value): self.embed_tokens = value def forward( self, input_ids: torch.LongTensor = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_values: Optional[Union[List[torch.FloatTensor], AttentionDynamicCache]] = None, inputs_embeds: Optional[torch.FloatTensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, fla_past_key_values = None, mamba_inference_params = None, ) -> Union[Tuple, MoeModelOutputWithPast]: 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 ) 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 input_ids is not None and inputs_embeds is not None: raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time") elif input_ids is not None: batch_size, seq_length = input_ids.shape elif inputs_embeds is not None: batch_size, seq_length, _ = inputs_embeds.shape else: raise ValueError("You have to specify either input_ids or inputs_embeds") if self.gradient_checkpointing and self.training: if use_cache: logger.warning_once( "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..." ) use_cache = False if position_ids is None: device = input_ids.device if input_ids is not None else inputs_embeds.device position_ids = torch.arange(0, seq_length, dtype=torch.long, device=device ) position_ids = position_ids.unsqueeze(0).view(-1, seq_length) else: if self.config.num_memory_tokens > 0 and past_key_values is not None and not self.has_previous_state: position_ids = position_ids.view(-1, seq_length + self.config.num_memory_tokens).long() else: position_ids = position_ids.view(-1, seq_length).long() if inputs_embeds is None: inputs_embeds = self.embed_tokens(input_ids) ori_b, ori_n = inputs_embeds.shape[0], inputs_embeds.shape[1] if self.config.num_memory_tokens > 0 and (past_key_values is None or not self.has_previous_state): mem = repeat(self.memory_tokens, 'n d -> b n d', b = inputs_embeds.shape[0]) # prepend the memory to every segment of m by repeating the memory tokens inputs_embeds, mem_packed_shape = pack((mem, inputs_embeds), 'b * d') if position_ids is not None and position_ids.shape[1] != inputs_embeds.shape[1]: position_ids = torch.arange(inputs_embeds.shape[1], device=inputs_embeds.device).unsqueeze(0) if attention_mask is not None and attention_mask.shape[1] < inputs_embeds.shape[1]: assert attention_mask.shape[1] + self.config.num_memory_tokens == inputs_embeds.shape[1] attention_mask = torch.cat([torch.ones(inputs_embeds.shape[0], self.config.num_memory_tokens, device=attention_mask.device), attention_mask], dim=1) if attention_mask is not None and self._attn_implementation == "flash_attention_2" and use_cache: is_padding_right = attention_mask[:, -1].sum().item() != batch_size if is_padding_right: raise ValueError( "You are attempting to perform batched generation with padding_side='right'" " this may lead to unexpected behaviour for Flash Attention version of NemotronFlash. Make sure to " " call `tokenizer.padding_side = 'left'` before tokenizing the input. " ) attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None hidden_states = inputs_embeds all_hidden_states = () if output_hidden_states else None all_self_attns = () if output_attentions else None next_decoder_cache = None for i, decoder_layer in enumerate(self.layers): if output_hidden_states: all_hidden_states += (hidden_states,) if self.gradient_checkpointing and self.training: layer_outputs = self._gradient_checkpointing_func( decoder_layer.__call__, hidden_states, attention_mask, position_ids, past_key_values, output_attentions, use_cache, ) else: layer_outputs = decoder_layer( hidden_states, attention_mask=attention_mask, position_ids=position_ids, past_key_value=past_key_values, output_attentions=output_attentions, use_cache=use_cache, use_swa=self.sliding_window is not None and i not in self.global_attn_idx, fla_past_key_values=fla_past_key_values, mamba_inference_params=mamba_inference_params, ) hidden_states = layer_outputs[0] if use_cache: next_decoder_cache = layer_outputs[2 if output_attentions else 1] if output_attentions: all_self_attns += (layer_outputs[1],) if self.final_layernorm is not None: hidden_states = self.final_layernorm(hidden_states) if output_hidden_states: all_hidden_states += (hidden_states,) if self.config.num_memory_tokens > 0 and (past_key_values is None or not self.has_previous_state): mem, hidden_states = unpack(hidden_states, mem_packed_shape, 'b * d') hidden_states = hidden_states[:, :ori_n, :] if past_key_values is not None and not self.has_previous_state: self.has_previous_state = True next_cache = None if use_cache: next_cache = next_decoder_cache if not return_dict: return tuple( v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None ) return MoeModelOutputWithPast( last_hidden_state=hidden_states, past_key_values=past_key_values if (fla_past_key_values is None and mamba_inference_params is None) else (past_key_values, fla_past_key_values, mamba_inference_params), hidden_states=all_hidden_states, attentions=all_self_attns, ) # Adapted from transformers.models.mixtral.modeling_mixtral.MixtralForCausalLM with MIXTRAL->JAMBA, Mixtral->NemotronFlash class NemotronFlashForCausalLM(NemotronFlashPreTrainedModel, GenerationMixin): _tied_weights_keys = ["lm_head.weight"] def __init__(self, config: NemotronFlashConfig): super().__init__(config) self.config = config self.model = NemotronFlashModel(config) self.vocab_size = config.vocab_size self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) 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 @replace_return_docstrings(output_type=MoeCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC) def forward( self, input_ids: torch.LongTensor = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_values: Optional[List[torch.FloatTensor]] = None, inputs_embeds: Optional[torch.FloatTensor] = None, labels: Optional[torch.LongTensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, calc_logits_for_entire_prompt: Optional[bool] = True, fla_past_key_values = None, mamba_inference_params = None, ) -> Union[Tuple, MoeCausalLMOutputWithPast]: r""" Args: labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): Labels for computing the masked language modeling loss. Indices should either be in `[0, ..., config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`. calc_logits_for_entire_prompt (`bool`, *optional*): Whether or not to calculate the logits for the entire prompt, or just the last token. Only last token logits are needed for generation, and calculating them only for that token can save memory, which becomes pretty significant for long sequences. Returns: ```""" 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.use_return_dict outputs = self.model( input_ids=input_ids, attention_mask=attention_mask, position_ids=position_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, fla_past_key_values=fla_past_key_values, mamba_inference_params=mamba_inference_params, return_dict=return_dict, ) hidden_states = outputs[0] if calc_logits_for_entire_prompt: logits = self.lm_head(hidden_states) else: logits = self.lm_head(hidden_states[..., -1:, :]) logits = logits / self.lm_head.weight.norm(p=2, dim=1) logits = logits.float() loss = None if labels is not None: shift_logits = logits[..., :-1, :].contiguous() shift_labels = labels[..., 1:].contiguous() loss_fct = CrossEntropyLoss() shift_logits = shift_logits.view(-1, self.config.vocab_size) shift_labels = shift_labels.view(-1) # Enable model parallelism shift_labels = shift_labels.to(shift_logits.device) loss = loss_fct(shift_logits, shift_labels) if not return_dict: output = (logits,) + outputs[1:] return (loss,) + output if loss is not None else output return MoeCausalLMOutputWithPast( loss=loss, logits=logits, past_key_values=outputs.past_key_values, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) def get_init_cache(self, max_seqlen, batch_size=1): past_key_values = AttentionDynamicCache( self.config, batch_size, self.dtype, device=self.device, layer_type=self.config.layer_type ) mamba_inference_params = InferenceParams(max_seqlen=max_seqlen, max_batch_size=batch_size) fla_past_key_values = fla_cache.from_legacy_cache(None) return past_key_values, fla_past_key_values, mamba_inference_params def init_cuda_graph_generation( self, max_new_tokens=128, batch_size=1, device=None, ): """ Initialize CUDA graph for generation with proper cache handling and warmup. This function should be called once before generation to set up the graph. Args: max_new_tokens: Maximum number of new tokens to generate batch_size: Batch size for generation device: Device to use (defaults to model device) Returns: generation_state: Dictionary containing all necessary state for generation """ if device is None: device = next(self.parameters()).device self.eval() # Initialize caches max_seqlen = max_new_tokens + 2048 + self.config.num_memory_tokens # Add buffer for input past_key_values, fla_past_key_values, mamba_inference_params = self.get_init_cache( max_seqlen=max_seqlen, batch_size=batch_size ) # Initialize KV caches for all modules for module in self.modules(): if hasattr(module, 'init_kv_cache'): module.init_kv_cache(max_batch_size=batch_size, max_seq_len=max_seqlen) with torch.no_grad(): # Warmup runs dummy_input = torch.ones((batch_size, 10), dtype=torch.long, device=device) for _ in range(10): self(dummy_input) # Prepare static tensors for CUDA graph static_current_input = torch.zeros((batch_size, 1), dtype=torch.long, device=device) static_position_ids = torch.zeros((batch_size, 1), dtype=torch.long, device=device) static_logits = torch.zeros((batch_size, self.config.vocab_size), device=device) # Set up for graph capture self.model.has_previous_state = True if mamba_inference_params is not None: mamba_inference_params.seqlen_offset = 1 # Warmup runs for graph capture for _ in range(10): model_kwargs_warmup = { 'input_ids': static_current_input, 'fla_past_key_values': fla_past_key_values, 'mamba_inference_params': mamba_inference_params, 'past_key_values': past_key_values, 'use_cache': True, 'position_ids': static_position_ids, } warmup_outputs = self(**model_kwargs_warmup) # Capture CUDA graph generation_graph = CUDAGraph() with torch.cuda.graph(generation_graph): model_kwargs_graph = { 'input_ids': static_current_input, 'fla_past_key_values': fla_past_key_values, 'mamba_inference_params': mamba_inference_params, 'past_key_values': past_key_values, 'use_cache': True, 'position_ids': static_position_ids, } graph_outputs = self(**model_kwargs_graph) static_logits.copy_(graph_outputs.logits[:, -1, :]) if fla_past_key_values is not None: fla_past_key_values.reset() if mamba_inference_params is not None: mamba_inference_params.reset(mamba_inference_params.max_seqlen, mamba_inference_params.max_batch_size) for key in mamba_inference_params.key_value_memory_dict: conv_state, ssm_state = mamba_inference_params.key_value_memory_dict[key] conv_state.zero_() ssm_state.zero_() for module in self.modules(): if hasattr(module, 'reset_kv_cache'): module.reset_kv_cache() self.model.has_previous_state = False # Return generation state generation_state = { 'generation_graph': generation_graph, 'static_current_input': static_current_input, 'static_position_ids': static_position_ids, 'static_logits': static_logits, 'past_key_values': past_key_values, 'fla_past_key_values': fla_past_key_values, 'mamba_inference_params': mamba_inference_params, 'max_seqlen': max_seqlen, 'batch_size': batch_size, 'device': device, } return generation_state def generate_with_cuda_graph( self, input_ids, generation_state, max_new_tokens=128, temperature=1.0, top_k=0, top_p=0.9, eos_token_id=None, verbose=False, profiling=False, ): """ Generate text using pre-initialized CUDA graph state. Args: input_ids: Input token IDs tensor of shape (batch_size, seq_len) generation_state: State dictionary returned by init_cuda_graph_generation max_new_tokens: Maximum number of new tokens to generate temperature: Sampling temperature (0 for greedy) top_k: Top-k filtering (0 to disable) top_p: Top-p filtering (1.0 to disable) eos_token_id: End-of-sequence token ID pad_token_id: Padding token ID verbose: Whether to print generated tokens profiling: Whether to return timing information Returns: generated_ids: Tensor of shape (batch_size, input_len + generated_len) or decode_latency if profiling=True """ self.eval() batch_size = input_ids.shape[0] device = input_ids.device # Extract state generation_graph = generation_state['generation_graph'] static_current_input = generation_state['static_current_input'] static_position_ids = generation_state['static_position_ids'] static_logits = generation_state['static_logits'] past_key_values = generation_state['past_key_values'] fla_past_key_values = generation_state['fla_past_key_values'] mamba_inference_params = generation_state['mamba_inference_params'] with torch.no_grad(): if mamba_inference_params.seqlen_offset == 0: if fla_past_key_values is not None: fla_past_key_values.reset() if mamba_inference_params is not None: mamba_inference_params.reset(mamba_inference_params.max_seqlen, mamba_inference_params.max_batch_size) for key in mamba_inference_params.key_value_memory_dict: conv_state, ssm_state = mamba_inference_params.key_value_memory_dict[key] conv_state.zero_() ssm_state.zero_() for module in self.modules(): if hasattr(module, 'reset_kv_cache'): module.reset_kv_cache() self.model.has_previous_state = False # Prefill phase - process input sequence position_ids = torch.arange( self.config.num_memory_tokens + input_ids.shape[1], dtype=torch.long, device=device ).unsqueeze(0).expand(batch_size, -1) else: # Prefill phase - process input sequence position_ids = torch.arange( mamba_inference_params.seqlen_offset, mamba_inference_params.seqlen_offset + input_ids.shape[1], dtype=torch.long, device=device ).unsqueeze(0).expand(batch_size, -1) current_input = input_ids model_kwargs = { 'input_ids': current_input, 'past_key_values': past_key_values, 'fla_past_key_values': fla_past_key_values, 'mamba_inference_params': mamba_inference_params, 'use_cache': True, 'position_ids': position_ids, } if profiling: torch.cuda.synchronize() t1 = time.time() # Forward pass for prefill outputs = self(**model_kwargs) if mamba_inference_params is not None: if mamba_inference_params.seqlen_offset == 0: mamba_inference_params.seqlen_offset = current_input.shape[1] + self.config.num_memory_tokens else: mamba_inference_params.seqlen_offset += current_input.shape[1] static_position_ids.fill_(position_ids[0, -1]) logits = outputs.logits[:, -1, :] # (batch_size, vocab_size) generated_tokens = [] # Generation loop using CUDA graph replay for step in range(max_new_tokens): # Sample next token using current logits if temperature == 0: next_token = torch.argmax(logits, dim=-1, keepdim=True) else: next_token = sample_token(logits, temperature=temperature, top_k=top_k, top_p=top_p) generated_tokens.append(next_token) # Check for EOS if not profiling and eos_token_id is not None and (next_token == eos_token_id).all(): if verbose: print("\nEOS reached") break # Update static tensors for graph replay static_current_input.copy_(next_token) static_position_ids.add_(1) # Replay the captured graph generation_graph.replay() if mamba_inference_params is not None: mamba_inference_params.seqlen_offset += 1 logits = static_logits.clone() generated_ids = torch.cat([input_ids] + generated_tokens, dim=1) if profiling: torch.cuda.synchronize() t2 = time.time() decode_latency = t2 - t1 return generated_ids, decode_latency return generated_ids def generate_with_cache( self, input_ids, max_new_tokens=128, temperature=1.0, top_k=0, top_p=0.9, eos_token_id=None, verbose=False, ): """ Generate text using the hybrid model with proper cache handling using pre-initialized CUDA graph state. Args: input_ids: Input token IDs tensor of shape (batch_size, seq_len) max_new_tokens: Maximum number of new tokens to generate temperature: Sampling temperature (0 for greedy) top_k: Top-k filtering (0 to disable) top_p: Top-p filtering (1.0 to disable) eos_token_id: End-of-sequence token ID verbose: Whether to print generated tokens Returns: generated_ids: Tensor of shape (batch_size, input_len + generated_len) """ self.eval() batch_size = input_ids.shape[0] device = input_ids.device with torch.no_grad(): max_seqlen = input_ids.shape[1] + max_new_tokens + self.config.num_memory_tokens past_key_values, fla_past_key_values, mamba_inference_params = self.get_init_cache(max_seqlen=max_seqlen, batch_size=batch_size) for module in self.model.modules(): if hasattr(module, 'init_kv_cache'): module.init_kv_cache(max_batch_size=batch_size, max_seq_len=max_seqlen) # Prefill phase - process input sequence current_input = input_ids position_ids = torch.arange( self.model.config.num_memory_tokens + current_input.shape[1], dtype=torch.long, device=device ).unsqueeze(0).expand(batch_size, -1) model_kwargs = { 'input_ids': current_input, 'past_key_values': past_key_values, 'fla_past_key_values': fla_past_key_values, 'mamba_inference_params': mamba_inference_params, 'use_cache': True, 'position_ids': position_ids, } outputs = self(**model_kwargs) # past_key_values, fla_past_key_values, mamba_inference_params = outputs.past_key_values mamba_inference_params.seqlen_offset = current_input.shape[1] + self.model.config.num_memory_tokens logits = outputs.logits[:, -1, :] # (batch_size, vocab_size) generated_tokens = [] # Generation loop for step in range(max_new_tokens): # Sample next token if temperature == 0: next_token = torch.argmax(logits, dim=-1, keepdim=True) else: next_token = sample_token(logits, temperature=temperature, top_k=top_k, top_p=top_p) generated_tokens.append(next_token) # Check for EOS if eos_token_id is not None and (next_token == eos_token_id).all(): if verbose: print("\nEOS reached") break current_input = next_token # Shape: (batch_size, 1) # Update position_ids for decoding if position_ids is not None: position_ids = torch.full( (batch_size, 1), position_ids[0, -1] + 1, dtype=torch.long, device=device ) # Forward pass for next token model_kwargs = { 'input_ids': current_input, 'fla_past_key_values': fla_past_key_values, 'mamba_inference_params': mamba_inference_params, 'past_key_values': past_key_values, 'use_cache': True, 'position_ids': position_ids, } outputs = self(**model_kwargs) mamba_inference_params.seqlen_offset += 1 logits = outputs.logits[:, -1, :] generated_ids = torch.cat([input_ids] + generated_tokens, dim=1) return generated_ids def prepare_inputs_for_generation( self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs, ): if self.config.num_memory_tokens > 0: attention_mask = torch.cat([torch.ones(input_ids.shape[0], self.config.num_memory_tokens, device=attention_mask.device), attention_mask], dim=1) ### Note that KV cache is disable when using model.generate; Please use model.generate_with_cuda_graph or model.generate_with_cache instead. past_key_values = None position_ids = kwargs.get("position_ids", None) if attention_mask is not None and position_ids is None: # create position_ids on the fly for batch generation position_ids = attention_mask.long().cumsum(-1) - 1 position_ids.masked_fill_(attention_mask == 0, 1) position_ids = position_ids[:, -input_ids.shape[1]:] # if `inputs_embeds` are passed, we only want to use them in the 1st generation step if inputs_embeds is not None: if input_ids.shape[1] == 0: model_inputs = {"inputs_embeds": inputs_embeds} else: inputs_embeds_new = self.model.embed_tokens(input_ids) model_inputs = {"inputs_embeds": torch.cat([inputs_embeds, inputs_embeds_new], dim=1)} else: model_inputs = {"input_ids": input_ids} model_inputs.update( { "position_ids": position_ids, "past_key_values": past_key_values, "use_cache": kwargs.get("use_cache"), "attention_mask": attention_mask, } ) return model_inputs def sample_token(logits, temperature=1.0, top_k=0, top_p=0.9): """ Sample a token from logits with temperature, top-k, and top-p filtering. Args: logits: Tensor of shape (batch_size, vocab_size) temperature: Sampling temperature top_k: Top-k filtering (0 to disable) top_p: Top-p filtering (1.0 to disable) Returns: next_token: Tensor of shape (batch_size, 1) """ if temperature == 0: return torch.argmax(logits, dim=-1, keepdim=True) logits = logits / temperature if top_k > 0: indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None] logits.masked_fill_(indices_to_remove, float('-inf')) if top_p < 1.0: sorted_logits, sorted_indices = torch.sort(logits, descending=True, dim=-1) cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1) # Remove tokens with cumulative probability above the threshold sorted_indices_to_remove = cumulative_probs > top_p # Shift the indices to the right to keep also the first token above the threshold sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone() sorted_indices_to_remove[..., 0] = 0 indices_to_remove = sorted_indices_to_remove.scatter(-1, sorted_indices, sorted_indices_to_remove) logits.masked_fill_(indices_to_remove, float('-inf')) probs = F.softmax(logits, dim=-1) return torch.multinomial(probs, num_samples=1)