import copy from functools import partial from typing import List, Optional, Tuple, Union import torch from torch.distributed.distributed_c10d import ProcessGroup from transformers.activations import ACT2FN from transformers.cache_utils import Cache, DynamicCache, StaticCache from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS from transformers.models.llama.modeling_llama import ( KwargsForCausalLM, LlamaAttention, LlamaDecoderLayer, LlamaForCausalLM, LlamaMLP, LlamaModel, LlamaRMSNorm, apply_rotary_pos_emb, logger, ) from transformers.processing_utils import Unpack from .... import distributed from ....distributed import parallel_state from ....distributed.strategy import DistributedStrategy, NoOpStrategy from ....distributed.tensorparallel import ( reduce_from_tensor_model_parallel_region, ) from ....distributed.tp import TPModule from ...modeling_attn_mask_utils import ( _gaudi_prepare_4d_causal_attention_mask, ) from ..modeling_all_models import Matmul, apply_customized_rope_module from .configuration_llama import LlamaConfig try: from habana_frameworks.torch.hpex.kernels import RotaryPosEmbeddingHelperV2 as FusedRoPE # noqa has_fused_rope = True except ImportError: has_fused_rope = False print("Not using HPU fused kernel for apply_rotary_pos_emb") try: from habana_frameworks.torch.hpex.normalization import FusedRMSNorm as FusedRMSNorm has_fused_rms_norm = True except ImportError: has_fused_rms_norm = False print("Not using HPU fused kernel for RMSNorm") try: from habana_frameworks.torch.hpex.kernels import FusedSDPA except ImportError: print("Not using HPU fused scaled dot-product attention kernel.") FusedSDPA = None import habana_frameworks.torch.core as htcore def gaudi_llama_rmsnorm_forward(self, hidden_states): """ Copied from LlamaRMSNorm.forward: https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/modeling_llama.py The only differences are: - override RMSNorm with Habana fused RMSNorm """ if hidden_states.device.type == "hpu" and has_fused_rms_norm: # mixed dtypes are not good for FusedRMSNorm, both inputs need to have same dtype if hidden_states.dtype != self.weight.dtype: orig_dtype = hidden_states.dtype hidden_states = FusedRMSNorm.apply(hidden_states.to(self.weight.dtype), self.weight, self.variance_epsilon) return hidden_states.to(orig_dtype) else: hidden_states = FusedRMSNorm.apply(hidden_states, self.weight, self.variance_epsilon) return hidden_states else: input_dtype = hidden_states.dtype hidden_states = hidden_states.to(torch.float32) variance = hidden_states.pow(2).mean(-1, keepdim=True) hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon) return self.weight * hidden_states.to(input_dtype) class GaudiLlamaRotaryEmbedding(torch.nn.Module): def __init__(self, config: LlamaConfig, device=None): super().__init__() # BC: "rope_type" was originally "type" 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 if self.rope_type == "linear": self.scaling_factor = config.rope_scaling["factor"] elif self.rope_type == "dynamic": self.scaling_factor = config.rope_scaling["factor"] self.base = config.rope_theta partial_rotary_factor = config.partial_rotary_factor if hasattr(config, "partial_rotary_factor") else 1.0 head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads) self.dim = int(head_dim * partial_rotary_factor) 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 # Build here to make `torch.jit.trace` work. self._set_cos_sin_cache( seq_len=self.max_seq_len_cached, device=self.inv_freq.device, dtype=torch.get_default_dtype() ) def _set_cos_sin_cache(self, seq_len, device, dtype): self.max_seq_len_cached = seq_len if self.rope_type == "dynamic" and seq_len > self.config.max_position_embeddings: base = self.base * ( (self.scaling_factor * seq_len / self.config.max_position_embeddings) - (self.scaling_factor - 1) ) ** (self.dim / (self.dim - 2)) inv_freq = 1.0 / (base ** (torch.arange(0, self.dim, 2).float().to(device) / self.dim)) self.register_buffer("inv_freq", inv_freq, persistent=False) # Use torch.int32 to avoid loss due to low precision with BF16 (refer to SW-215204) t = torch.arange(self.max_seq_len_cached, device=device, dtype=torch.int32) if self.rope_type == "linear": t = t / self.scaling_factor freqs = torch.outer(t, self.inv_freq) # Different from paper, but it uses a different permutation in order to obtain the same calculation emb = torch.cat((freqs, freqs), dim=-1) self.register_buffer("_cos_cached", emb.cos().to(dtype), persistent=False) self.register_buffer("_sin_cached", emb.sin().to(dtype), persistent=False) def _dynamic_frequency_update(self, seq_len, 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 inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device, seq_len=seq_len) self.register_buffer("inv_freq", inv_freq, persistent=False) # TODO joao: may break with compilation self.max_seq_len_cached = seq_len if seq_len < self.original_max_seq_len and self.max_seq_len_cached > self.original_max_seq_len: # reset # This .to() is needed if the model has been moved to a device after being initialized (because # the buffer is automatically moved, but not the original copy) self.original_inv_freq = self.original_inv_freq.to(device) self.register_buffer("inv_freq", self.original_inv_freq, persistent=False) self.max_seq_len_cached = self.original_max_seq_len @torch.no_grad() def forward(self, x, seq_len=None): # x: [bs, num_attention_heads, seq_len, head_size] if "dynamic" in self.rope_type: self._dynamic_frequency_update(seq_len, device=x.device) if seq_len > self.max_seq_len_cached: self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype) if self.attention_scaling == 1.0: return ( self._cos_cached[:seq_len].to(dtype=x.dtype), self._sin_cached[:seq_len].to(dtype=x.dtype), ) else: return ( self._cos_cached[:seq_len].to(dtype=x.dtype) * self.attention_scaling, self._sin_cached[:seq_len].to(dtype=x.dtype) * self.attention_scaling, ) class GaudiLlamaMLP(LlamaMLP): def __init__(self, config): super(LlamaMLP, self).__init__() self.config = config self.hidden_size = config.hidden_size self.intermediate_size = config.intermediate_size self.gate_proj = torch.nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias) self.up_proj = torch.nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias) self.down_proj = torch.nn.Linear(self.intermediate_size, self.hidden_size, bias=config.mlp_bias) self.act_fn = ACT2FN[config.hidden_act] def pre_mlp_forward(self, x): input = self.act_fn(self.gate_proj(x)) * self.up_proj(x) output = self.down_proj(input) return output def mlp_all_reduce(self, x): if hasattr(self.down_proj, "all_reduce"): self.down_proj.all_reduce(x) def post_mlp_forward(self, x): if hasattr(self.down_proj, "post_all_reduce"): return self.down_proj.post_all_reduce(x) return x class TPGaudiLlamaMLP(GaudiLlamaMLP, TPModule): def __init__( self, config, group: Optional[ProcessGroup] = None, ): assert torch.distributed.is_initialized() rank, world_size = distributed.rank_and_world(group) hidden_dim = int(config.hidden_grow_factor * config.hidden_size) assert hidden_dim % world_size == 0, "Hidden dim must be divisible by world size" self.config = copy.deepcopy(config) self.config.intermediate_size = int((config.hidden_grow_factor / world_size) * config.hidden_size) GaudiLlamaMLP.__init__(self, self.config) self.setup_tp(rank, world_size) def colwise_param_names(self) -> List[str]: return ["up_proj", "gate_proj"] def rowwise_param_names(self) -> List[str]: return ["down_proj"] @staticmethod def import_module(glu: GaudiLlamaMLP, group: ProcessGroup) -> "TPGaudiLlamaMLP": config = copy.deepcopy(glu.config) config.hidden_grow_factor = glu.config.intermediate_size / glu.config.hidden_size tp_glu = TPGaudiLlamaMLP(config=config, group=group) return tp_glu def pre_mlp_forward(self, x): out_par = GaudiLlamaMLP.pre_mlp_forward(self, x) return reduce_from_tensor_model_parallel_region(out_par) def gaudi_llama_repeat_kv( query_states: torch.Tensor, key_states: torch.Tensor, value_states: torch.Tensor, attention_mask: torch.Tensor, n_rep: int, ): """ Copied from repeat_kv: https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/modeling_llama.py The only differences are: - Append num_key_value_heads == 1 check as kv states can be broadcasted during matmuls so need to expand and reshape them. - Add new args query_states, key_states, value_states and attention_mask and update the logic for expansion. The query states go from (batch, num_heads, seqlen, head_dim) to (batch, num_key_value_heads, n_rep, seqlen, head_dim) The key/value states go from (batch, num_key_value_heads, seqlen, head_dim) to (batch, num_key_value_heads, 1, seqlen, head_dim) """ batch, num_key_value_heads, kv_len, head_dim = key_states.shape if n_rep == 1 or num_key_value_heads == 1: return query_states, key_states, value_states, attention_mask new_kv_shape = (batch, num_key_value_heads, 1, kv_len, head_dim) key_states = key_states.reshape(new_kv_shape) value_states = value_states.reshape(new_kv_shape) batch, _, q_len, head_dim = query_states.shape new_q_shape = (batch, num_key_value_heads, n_rep, q_len, head_dim) query_states = query_states.reshape(new_q_shape) if attention_mask is not None: # Add groups dim and set to 1 attention_mask = attention_mask.unsqueeze(1) return query_states, key_states, value_states, attention_mask # FusedScaledDotProductAttention class ModuleFusedSDPA(torch.nn.Module): def __init__(self, fusedSDPA, scale, attention_dropout, enable_recompute, flash_attention_fp8): super().__init__() self._hpu_kernel_fsdpa = fusedSDPA self.scale = scale self.attention_dropout = attention_dropout self.enable_recompute = enable_recompute self.flash_attention_fp8 = flash_attention_fp8 def forward( self, query, key, value, attn_mask, dropout_p, is_causal, scale, softmax_mode, recompute_mode, valid_sequence_lengths, padding_side="left", ): return self._hpu_kernel_fsdpa.apply( query, key, value, attn_mask, dropout_p, is_causal, scale, softmax_mode, recompute_mode, valid_sequence_lengths, padding_side, ) class KVCache(torch.nn.Module): def __init__(self): super(KVCache, self).__init__() self.cache = None self.inp_seq_len = -1 def allocate(self, inp_seq_len, dtype, device, shape): if self.cache is None or self.cache.shape != shape: self.inp_seq_len = inp_seq_len self.cache = torch.zeros(shape, dtype=dtype, device=device) else: assert self.inp_seq_len == inp_seq_len, ( f"inp_seq_len must be the same. self.inp_seq_len:{self.inp_seq_len} inp_seq_len:{inp_seq_len}" ) self.cache.fill_(0) @staticmethod def update(prev, cur, dim, idx, inp_seq_len): if inp_seq_len != -1: # reuse cache logic orig_cur = cur if prev.shape == cur.shape: prev.copy_(cur) return orig_cur if cur.shape[2] > 1 and cur.shape[2] <= prev.shape[2]: # Initialize prev[:, :, :inp_seq_len, :].copy_(cur) return orig_cur if idx is not None: # 2+ tokenizer logic if model is static shape optimized prev.index_copy_(dim, idx - 1, cur) return prev else: return torch.cat((prev, cur), dim=dim) def get_shape(self): if self.cache is None: return None return self.cache.shape def forward(self, cur, dim, idx): return self.update(self.cache, cur, dim, idx, self.inp_seq_len) class GaudiDistributedAttention(torch.nn.Module): def __init__( self, hpu_module_fsdpa: ModuleFusedSDPA, scale, attention_dropout, enable_recompute, flash_attention_fp8 ): super().__init__() self._hpu_module_fsdpa = hpu_module_fsdpa if parallel_state.sequence_parallel_is_initialized() and parallel_state.get_sequence_parallel_world_size() > 1: from deepspeed.sequence.layer import DistributedAttention self._hpu_module_fsdpa_distributed = DistributedAttention( self._hpu_module_fsdpa, parallel_state.get_sequence_parallel_group(), 1, 2 ) def forward( self, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, attn_mask: torch.Tensor, dropout_p: float, is_casual, scale, softmax_mode, recompute_mode, valid_sequence_lengths, padding_side="left", ): if parallel_state.sequence_parallel_is_initialized() and parallel_state.get_sequence_parallel_world_size() > 1: return self._hpu_module_fsdpa_distributed( query, key, value, 0, # As the shape for inputs is [B, N, S, H] None, attn_mask, dropout_p, is_casual, scale, softmax_mode, recompute_mode, valid_sequence_lengths, padding_side, ) else: return self._hpu_module_fsdpa( query, key, value, attn_mask, dropout_p, is_casual, scale, softmax_mode, recompute_mode, valid_sequence_lengths, padding_side, ) def get_gaudi_distributed_attention( fused_scaled_dot_product_attention, fused_scaled_dot_product_attention_distributed ): if parallel_state.sequence_parallel_is_initialized() and parallel_state.get_sequence_parallel_world_size() > 1: return fused_scaled_dot_product_attention_distributed else: return fused_scaled_dot_product_attention def gaudi_eager_attention_forward( module: torch.nn.Module, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, attention_mask: Optional[torch.Tensor], scaling: float, dropout: float = 0.0, attn_softmax_bf16: bool = False, **kwargs, ): bsz, q_len = kwargs["input_shape"] query_states, key_states, value_states, attention_mask = gaudi_llama_repeat_kv( query, key, value, attention_mask, module.num_key_value_groups ) attn_weights = module.matmul_qk(query_states, key_states.transpose(-2, -1)) * scaling if attention_mask is not None: causal_mask = attention_mask[:, :, :, : key_states.shape[-2]] attn_weights = attn_weights + causal_mask if attn_softmax_bf16: attn_weights = torch.nn.functional.softmax(attn_weights, dim=-1, dtype=query_states.dtype) else: # upcast attention to fp32 attn_weights = torch.nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype) attn_weights = torch.nn.functional.dropout(attn_weights, p=dropout, training=module.training) attn_output = module.matmul_av(attn_weights, value_states) attn_output = attn_output.reshape(bsz, -1, q_len, module.head_dim) return attn_output, attn_weights class GaudiLlamaAttention(LlamaAttention): def __init__(self, config: LlamaConfig, layer_idx: Optional[int] = None): super().__init__(config, layer_idx) self.matmul_qk = Matmul() self.matmul_av = Matmul() self.k_cache = KVCache() self.v_cache = KVCache() self.rotary_emb = GaudiLlamaRotaryEmbedding(config=config) self.num_key_value_heads = config.num_key_value_heads self.hidden_size = config.hidden_size self.num_heads = config.num_attention_heads self.head_dim = getattr(config, "head_dim", self.hidden_size // self.num_heads) if hasattr(config, "fused_qkv") and config.fused_qkv: self.num_heads = config.num_attention_heads self.head_dim = config.hidden_size // self.num_heads self.dim1 = self.num_heads * self.head_dim self.dim2 = config.num_key_value_heads * self.head_dim self.qkv_proj = torch.nn.Linear( self.hidden_size, self.dim1 + 2 * self.dim2, bias=config.attention_bias, ) self.q_proj = None self.k_proj = None self.v_proj = None self.inp_seq_len = -1 self.fused_scaled_dot_product_attention = ( ModuleFusedSDPA( FusedSDPA, scale=self.scaling, attention_dropout=self.attention_dropout, enable_recompute=False, flash_attention_fp8=getattr(config, "flash_attention_fp8", False), ) if FusedSDPA else None ) # for all2all comm, Distributed Attention cares about sequence (s) and number of heads (h) dimensions. In HPU, they are at 1 and 2 indices self.fused_scaled_dot_product_attention_distributed = None if parallel_state.sequence_parallel_is_initialized() and parallel_state.get_sequence_parallel_world_size() > 1: self.fused_scaled_dot_product_attention_distributed = ( GaudiDistributedAttention( self.fused_scaled_dot_product_attention, scale=self.scaling, attention_dropout=self.attention_dropout, enable_recompute=False, flash_attention_fp8=getattr(config, "flash_attention_fp8", False), ) if FusedSDPA else None ) def get_k_proj_weight(self): """4bit quantization in GPTQ replaces the k_proj.weight with qweight.""" if hasattr(self.k_proj, "qweight"): return self.k_proj.qweight return self.k_proj.weight def get_k_proj_weight_dtype(self): """4bit quantization in GPTQ replaces the k_proj.weight with qweight. Scales tensor gets the weight dtype.""" if hasattr(self.k_proj, "qweight"): return self.k_proj.scales.dtype elif hasattr(self.k_proj, "use_qdq") and self.k_proj.use_qdq: return self.k_proj.dequant_weights.hp_dtype return self.k_proj.weight.dtype def allocate_kv_cache(self, batch_size, max_seq_len, inp_seq_len): cache_shape = (batch_size, self.num_key_value_heads, max_seq_len, self.head_dim) device = self.get_k_proj_weight().device dtype = self.config.torch_dtype self.k_cache.allocate(inp_seq_len, dtype, device, cache_shape) self.v_cache.allocate(inp_seq_len, dtype, device, cache_shape) def update_sincos_cache(self, seq_len): # Call rotary emb forward() to update cos/sin cache when infering more than self.rotary_emb.original_max_seq_len # This helps in avoiding creation of these caches during actual model forward pass and # reduce memory consumption and improve performance. if seq_len > self.rotary_emb.original_max_seq_len: self.rotary_emb.original_max_seq_len = seq_len _, _ = self.rotary_emb(self.get_k_proj_weight(), seq_len=seq_len) def reorder(self, tensor, beam_idx, dim_a, dim_b): updated = tensor.index_select(0, beam_idx) tensor.copy_(updated) def reorder_kv_cache(self, beam_idx: torch.LongTensor): if self.k_cache.cache is None: return (None, None) head_dim = self.k_cache.cache.size(-1) seq_length = self.k_cache.cache.size(-2) self.reorder(self.k_cache.cache, beam_idx, seq_length, head_dim) self.reorder(self.v_cache.cache, beam_idx, seq_length, head_dim) return (self.k_cache.cache.shape, self.v_cache.cache.shape) def pre_attn_forward( self, hidden_states: torch.Tensor, position_embeddings: Tuple[torch.Tensor, torch.Tensor], attention_mask: Optional[torch.Tensor], past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, cache_position: Optional[torch.LongTensor] = None, position_ids: Optional[torch.LongTensor] = None, token_idx: Optional[torch.Tensor] = None, attn_softmax_bf16: Optional[bool] = False, reuse_cache: Optional[bool] = False, use_flash_attention: Optional[bool] = False, flash_attention_recompute: Optional[bool] = False, flash_attention_causal_mask: Optional[bool] = False, flash_attention_fast_softmax: Optional[bool] = False, valid_sequence_lengths: Optional[torch.Tensor] = None, cache_idx: int = None, num_virtual_tokens: int = None, **kwargs, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: """ Copied from LlamaAttention.forward: https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/modeling_llama.py The only differences are: - add new args token_idx - optimize KV cache - add new args attn_softmax_bf16 - add new args reuse_cache - add new args use_flash_attention - add new arg flash_attention_recompute - add new arg flash_attention_causal_mask - add new arg flash_attention_fast_softmax - add new arg num_virtual_tokens """ input_shape = hidden_states.shape[:-1] q_len = input_shape[1] hidden_shape = (*input_shape, -1, self.head_dim) if hasattr(self.config, "fused_qkv") and self.config.fused_qkv: qkv_states = self.qkv_proj(hidden_states) query_states, key_states, value_states = torch.split(qkv_states, [self.dim1, self.dim2, self.dim2], dim=-1) else: query_states = self.q_proj(hidden_states) key_states = self.k_proj(hidden_states) value_states = self.v_proj(hidden_states) # use -1 to infer num_heads and num_key_value_heads as they may vary if tensor parallel is used query_states = query_states.view(hidden_shape).transpose(1, 2) # TODO: update when auto mp params is enabled in DeepSpeed (cf. https://github.com/HabanaAI/DeepSpeed/blob/94309c7b5dfc1a69858f5c9f25737b2f81a332a5/deepspeed/module_inject/replace_module.py#L440) key_states = key_states.view(hidden_shape).transpose(1, 2) value_states = value_states.view(hidden_shape).transpose(1, 2) kv_seq_len = key_states.shape[-2] if past_key_value is not None: if token_idx is None: if hasattr(past_key_value, "get_usable_length"): kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx) else: kv_seq_len += past_key_value[0].shape[-2] else: if reuse_cache and not isinstance(past_key_value[0], torch.Tensor): kv_seq_len = past_key_value[0][-2] else: if num_virtual_tokens is not None and num_virtual_tokens == past_key_value[0].shape[-2]: kv_seq_len = past_key_value[0].shape[-2] + kv_seq_len else: kv_seq_len = past_key_value[0].shape[-2] # TODO: the following section cause torch.compile performance issue with graph recompilation # as we are not using position_embeddings, disable it for now # if position_embeddings is None: # logger.warning_once( # "The attention layers in this model are transitioning from computing the RoPE embeddings internally " # "through `position_ids` (2D tensor with the indexes of the tokens), to using externally computed " # "`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.46 `position_ids` will be " # "removed and `position_embeddings` will be mandatory." # ) # cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len) # else: # cos, sin = position_embeddings seq_len = kv_seq_len if parallel_state.sequence_parallel_is_initialized(): seq_len = kv_seq_len * parallel_state.get_sequence_parallel_world_size() cos, sin = self.rotary_emb(value_states, seq_len=seq_len) # If sequence parallel in enabled, position_ids should be based on which part of the sequence is present in the rank # As we divide the inputs based on ranks, position_ids are generated to suit that part of the sequence if parallel_state.sequence_parallel_is_initialized() and parallel_state.get_sequence_parallel_rank() > 0: position_ids = torch.arange( kv_seq_len * parallel_state.get_sequence_parallel_rank(), kv_seq_len * (parallel_state.get_sequence_parallel_rank() + 1), dtype=torch.long, device=query_states.device, ) position_ids = position_ids.unsqueeze(0) query_states, key_states = apply_customized_rope( query_states, key_states, cos, sin, position_ids, self.training ) if use_cache: # reuse k, v, self_attention if reuse_cache: if past_key_value is not None and isinstance(past_key_value[0], torch.Tensor): # prefix tuning case. attach past_key_value to generate first token. key_states = torch.cat((past_key_value[0], key_states), -2) value_states = torch.cat((past_key_value[1], value_states), -2) key_states = self.k_cache(key_states, 2, token_idx) value_states = self.v_cache(value_states, 2, token_idx) past_key_value = (self.k_cache.get_shape(), self.v_cache.get_shape()) else: if past_key_value is None: past_key = torch.zeros( key_states.shape, dtype=self.get_k_proj_weight_dtype() if self.get_k_proj_weight_dtype() != torch.uint8 else key_states.dtype, device=key_states.device, ) past_value = torch.zeros( key_states.shape, dtype=self.get_k_proj_weight_dtype() if self.get_k_proj_weight_dtype() != torch.uint8 else key_states.dtype, device=key_states.device, ) # Return list instead of tuple past_key_value = [past_key, past_value] key_states = self.k_cache.update(past_key_value[0], key_states, 2, token_idx, key_states.shape[-2]) value_states = self.v_cache.update( past_key_value[1], value_states, 2, token_idx, value_states.shape[-2] ) elif ( token_idx is not None and num_virtual_tokens is not None and num_virtual_tokens == past_key_value[0].shape[-2] ): # prefix tuning case. attach past_key_value to generate first token. key_states = self.k_cache.update(past_key_value[0], key_states, 2, None, -1) value_states = self.v_cache.update(past_key_value[1], value_states, 2, None, -1) past_key_value = (key_states, value_states) else: key_states = self.k_cache.update(past_key_value[0], key_states, 2, token_idx, self.inp_seq_len) value_states = self.v_cache.update(past_key_value[1], value_states, 2, token_idx, self.inp_seq_len) if token_idx is None: past_key_value = (key_states, value_states) if cache_idx is not None and q_len == 1: key_states = key_states[:, :, :cache_idx, :] value_states = value_states[:, :, :cache_idx, :] if attention_mask is not None: attention_mask = attention_mask[:, :, :, :cache_idx] kv_seq_len = key_states.shape[-2] else: past_key_value = None fused_scaled_dot_product_attention = get_gaudi_distributed_attention( self.fused_scaled_dot_product_attention, self.fused_scaled_dot_product_attention_distributed ) if use_flash_attention and FusedSDPA is not None: attn_weights = None if q_len == 1: # next token attn_output = fused_scaled_dot_product_attention( query_states, key_states, value_states, attention_mask, 0.0, False, None, "None", False, None, "None", ) else: # first token softmax_mode = "fast" if flash_attention_fast_softmax else "None" if flash_attention_causal_mask: # causal masking on first token requires inputs to be of the same length attn_output = fused_scaled_dot_product_attention( query_states, key_states, value_states, None, 0.0, True, None, softmax_mode, flash_attention_recompute, valid_sequence_lengths, "left", ) else: attn_output = fused_scaled_dot_product_attention( query_states, key_states, value_states, attention_mask, 0.0, False, None, softmax_mode, flash_attention_recompute, None, "None", ) else: attn_output, attn_weights = gaudi_eager_attention_forward( self, query_states, key_states, value_states, attention_mask, dropout=0.0 if not self.training else self.attention_dropout, scaling=self.scaling, attn_softmax_bf16=attn_softmax_bf16, input_shape=input_shape, ) attn_output = attn_output.transpose(1, 2).contiguous() attn_output = attn_output.reshape(*input_shape, -1).contiguous() attn_output = self.o_proj(attn_output) if not output_attentions: attn_weights = None if not reuse_cache and token_idx is not None and cache_idx is not None and q_len == 1: # Return only past key value shapes and not the tensors during decode phase (q len is 1) # to avoid making past key values as persistent output tensors of HPU graphs. past_key_value = (past_key_value[0].shape, past_key_value[1].shape) return attn_output, attn_weights, past_key_value def attention_all_reduce(self, attn_output): if hasattr(self.o_proj, "all_reduce"): self.o_proj.all_reduce(attn_output) def post_attn_forward(self, attn_output): if hasattr(self.o_proj, "post_all_reduce"): return self.o_proj.post_all_reduce(attn_output) return attn_output class TPGaudiLlamaAttention(GaudiLlamaAttention, TPModule): def __init__( self, config: LlamaConfig, layer_idx: Optional[int] = None, group: Optional[ProcessGroup] = None, ): super().__init__(config, layer_idx) assert torch.distributed.is_initialized() rank, world_size = distributed.rank_and_world(group) assert config.num_attention_heads % world_size == 0, "The number of heads must be divisible by world size" self.config = copy.deepcopy(config) self.pre_tp_kvheads = config.num_key_value_heads GaudiLlamaAttention.__init__(self, self.config, layer_idx) self.config.num_attention_heads = self.config.num_attention_heads // world_size self.config.num_key_value_heads = ( (self.config.num_key_value_heads // world_size) if self.config.num_key_value_heads > 1 else self.config.num_key_value_heads ) self.head_dim = config.hidden_size // config.num_attention_heads self.hidden_size = self.config.hidden_size // world_size self.num_heads = self.config.num_attention_heads self.q_proj = torch.nn.Linear( config.hidden_size, self.config.num_attention_heads * self.head_dim, bias=config.attention_bias ) self.k_proj = torch.nn.Linear( config.hidden_size, self.config.num_key_value_heads * self.head_dim, bias=config.attention_bias ) self.v_proj = torch.nn.Linear( config.hidden_size, self.config.num_key_value_heads * self.head_dim, bias=config.attention_bias ) self.o_proj = torch.nn.Linear( self.config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias ) self.setup_tp(rank, world_size) def colwise_param_names(self) -> List[str]: colwise_weights = ["q_proj"] if self.pre_tp_kvheads != 1: colwise_weights.append("k_proj") colwise_weights.append("v_proj") return colwise_weights def rowwise_param_names(self) -> List[str]: return ["o_proj"] @staticmethod def import_module(mha: GaudiLlamaAttention, layer_idx, group: ProcessGroup) -> "TPGaudiLlamaAttention": tp_mha = TPGaudiLlamaAttention(config=mha.config, layer_idx=layer_idx, group=group) return tp_mha def pre_attn_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, cache_position: Optional[torch.LongTensor] = None, position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, # will become mandatory in v4.45 token_idx: Optional[torch.Tensor] = None, attn_softmax_bf16: Optional[bool] = False, reuse_cache: Optional[bool] = False, use_flash_attention: Optional[bool] = False, flash_attention_recompute: Optional[bool] = False, flash_attention_causal_mask: Optional[bool] = False, flash_attention_fast_softmax: Optional[bool] = False, cache_idx: int = None, **kwargs, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: hidden_states, attn_weights, present_key_value = GaudiLlamaAttention.pre_attn_forward( self, 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, cache_position=cache_position, position_embeddings=position_embeddings, token_idx=token_idx, attn_softmax_bf16=attn_softmax_bf16, reuse_cache=reuse_cache, use_flash_attention=use_flash_attention, flash_attention_recompute=flash_attention_recompute, flash_attention_causal_mask=flash_attention_causal_mask, flash_attention_fast_softmax=flash_attention_fast_softmax, cache_idx=cache_idx, **kwargs, ) hidden_states = reduce_from_tensor_model_parallel_region(hidden_states) return hidden_states, attn_weights, present_key_value class GaudiLlamaDecoderLayer(LlamaDecoderLayer): def __init__(self, config: LlamaConfig, layer_idx: int): super(LlamaDecoderLayer, self).__init__() self.hidden_size = config.hidden_size self.self_attn = GaudiLlamaAttention(config=config, layer_idx=layer_idx) self.mlp = GaudiLlamaMLP(config) self.input_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.post_attention_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps) def allocate_kv_cache(self, batch_size, max_seq_len, inp_seq_len): self.self_attn.allocate_kv_cache(batch_size, max_seq_len, inp_seq_len) def reorder_kv_cache(self, beam_idx: torch.LongTensor): return self.self_attn.reorder_kv_cache(beam_idx) def update_sincos_cache(self, seq_len): self.self_attn.update_sincos_cache(seq_len) 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: Optional[bool] = False, use_cache: Optional[bool] = False, cache_position: Optional[torch.LongTensor] = None, position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, # necessary, but kept here for BC token_idx: Optional[torch.Tensor] = None, attn_softmax_bf16: Optional[bool] = False, reuse_cache: Optional[bool] = False, use_flash_attention: Optional[bool] = False, flash_attention_recompute: Optional[bool] = False, flash_attention_causal_mask: Optional[bool] = False, flash_attention_fast_softmax: Optional[bool] = False, valid_sequence_lengths: Optional[torch.Tensor] = None, cache_idx: int = None, num_virtual_tokens: int = None, attn_batch_split: int = 1, prev_layer_residual: Optional[torch.Tensor] = None, **kwargs, ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]: """ Copied from LlamaDecoderLayer.forward: https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/modeling_llama.py The only differences are: - add new args token_idx - add new args attn_softmax_bf16 - add new args reuse_cache - add new args use_flash_attention - add new arg flash_attention_recompute - add new arg flash_attention_causal_mask - add new arg flash_attention_fast_softmax """ if attn_batch_split > 1 and past_key_value is None: # Calculate split sizes to handle cases where batch size is not divisible by attn_batch_split batch_size = attention_mask.size(0) base_split_size = batch_size // attn_batch_split remainder = batch_size % attn_batch_split split_sizes = [base_split_size + 1 if i < remainder else base_split_size for i in range(attn_batch_split)] # Split tensors using the calculated sizes sub_attention_mask = torch.split(attention_mask, split_sizes, dim=0) sub_position_ids = torch.split(position_ids, split_sizes, dim=0) sub_valid_sequence_lengths = torch.split(valid_sequence_lengths, split_sizes, dim=0) split_attn_weights = [] split_present_key_values = [] split_hidden_states = [None] * attn_batch_split residual = [None] * attn_batch_split for i in range(attn_batch_split): split_hidden_states[i] = hidden_states[i] if self.self_attn.layer_idx != 0: # Add the residual from the previous layer split_hidden_states[i] = self.post_mlp(hidden_states[i], prev_layer_residual[i]) residual[i] = split_hidden_states[i] split_hidden_states[i], self_attn_weights, present_key_value = self.pre_attn( hidden_states=split_hidden_states[i], attention_mask=sub_attention_mask[i], position_ids=sub_position_ids[i], past_key_value=past_key_value, output_attentions=output_attentions, use_cache=use_cache, cache_position=cache_position, position_embeddings=position_embeddings, token_idx=token_idx, attn_softmax_bf16=attn_softmax_bf16, reuse_cache=reuse_cache, use_flash_attention=use_flash_attention, flash_attention_recompute=flash_attention_recompute, flash_attention_causal_mask=flash_attention_causal_mask, flash_attention_fast_softmax=flash_attention_fast_softmax, valid_sequence_lengths=sub_valid_sequence_lengths[i], cache_idx=cache_idx, num_virtual_tokens=num_virtual_tokens, **kwargs, ) self.self_attn.attention_all_reduce(split_hidden_states[i]) if output_attentions: split_attn_weights.append(self_attn_weights) if use_cache: split_present_key_values.append(present_key_value) self_attn_weights = torch.cat(split_attn_weights, dim=0) if split_attn_weights else None present_key_value = [torch.cat(tensors, dim=0) for tensors in zip(*split_present_key_values)] int_residual_splits = [] for i in range(attn_batch_split): split_hidden_states[i], int_residual = self.post_attn_pre_mlp(split_hidden_states[i], residual[i]) self.mlp.mlp_all_reduce(split_hidden_states[i]) int_residual_splits.append(int_residual) if self.self_attn.layer_idx == (self.self_attn.config.num_hidden_layers - 1): for i in range(attn_batch_split): split_hidden_states[i] = self.post_mlp(split_hidden_states[i], int_residual_splits[i]) hidden_states = split_hidden_states else: residual = hidden_states hidden_states, self_attn_weights, present_key_value = self.pre_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, cache_position=cache_position, position_embeddings=position_embeddings, token_idx=token_idx, attn_softmax_bf16=attn_softmax_bf16, reuse_cache=reuse_cache, use_flash_attention=use_flash_attention, flash_attention_recompute=flash_attention_recompute, flash_attention_causal_mask=flash_attention_causal_mask, flash_attention_fast_softmax=flash_attention_fast_softmax, valid_sequence_lengths=valid_sequence_lengths, cache_idx=cache_idx, num_virtual_tokens=num_virtual_tokens, **kwargs, ) self.self_attn.attention_all_reduce(hidden_states) hidden_states, residual = self.post_attn_pre_mlp(hidden_states, residual) self.mlp.mlp_all_reduce(hidden_states) hidden_states = self.post_mlp(hidden_states, residual) outputs = (hidden_states,) if output_attentions: outputs += (self_attn_weights,) if use_cache: outputs += (present_key_value,) # Store the residual splits to add them in the beginning of the next layer if attn_batch_split > 1 and past_key_value is None: outputs += (int_residual_splits,) return outputs def pre_attn( 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, cache_position: Optional[torch.LongTensor] = None, position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, # will become mandatory in v4.45 token_idx: Optional[torch.Tensor] = None, attn_softmax_bf16: Optional[bool] = False, reuse_cache: Optional[bool] = False, use_flash_attention: Optional[bool] = False, flash_attention_recompute: Optional[bool] = False, flash_attention_causal_mask: Optional[bool] = False, flash_attention_fast_softmax: Optional[bool] = False, valid_sequence_lengths: Optional[torch.Tensor] = None, cache_idx: int = None, num_virtual_tokens: int = None, ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]: hidden_states = self.input_layernorm(hidden_states) hidden_states, attn_weights, present_key_value = self.self_attn.pre_attn_forward( 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, cache_position=cache_position, position_embeddings=position_embeddings, token_idx=token_idx, attn_softmax_bf16=attn_softmax_bf16, reuse_cache=reuse_cache, use_flash_attention=use_flash_attention, flash_attention_recompute=flash_attention_recompute, flash_attention_causal_mask=flash_attention_causal_mask, flash_attention_fast_softmax=flash_attention_fast_softmax, valid_sequence_lengths=valid_sequence_lengths, cache_idx=cache_idx, num_virtual_tokens=num_virtual_tokens, ) return hidden_states, attn_weights, present_key_value def post_attn_pre_mlp(self, hidden_states, residual): hidden_states = self.self_attn.post_attn_forward(hidden_states) if self.training: hidden_states = hidden_states + residual residual = hidden_states else: residual.add_(hidden_states) hidden_states = residual hidden_states = self.post_attention_layernorm(hidden_states) hidden_states = self.mlp.pre_mlp_forward(hidden_states) return hidden_states, residual def post_mlp(self, hidden_states, residual): hidden_states = self.mlp.post_mlp_forward(hidden_states) if self.training: hidden_states = hidden_states + residual else: residual.add_(hidden_states) hidden_states = residual return hidden_states class GaudiLlamaModel(LlamaModel): """ Copied from https://github.com/huggingface/transformers/blob/v4.38.2/src/transformers/models/llama/modeling_llama.py#L909 """ def __init__(self, config: LlamaConfig): """ Copied from https://github.com/huggingface/transformers/blob/v4.38.2/src/transformers/models/llama/modeling_llama.py#L917 1. set fill_value to 1 instead of True 2. add device=self.device """ super(LlamaModel, self).__init__(config) self.padding_idx = config.pad_token_id self.vocab_size = config.vocab_size self.embed_tokens = torch.nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx) layers = [] for layer_idx in range(config.num_hidden_layers): layer = GaudiLlamaDecoderLayer(config, layer_idx) if hasattr(config, "parallel_strategy") and config.parallel_strategy is not None: layer = config.parallel_strategy.distribute_layer(layer, layer_idx) layers.append(layer) self.layers = torch.nn.ModuleList(layers) # parallel_strategy is not JSON serializable config.parallel_strategy = None self.norm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.gradient_checkpointing = False # Initialize weights and apply final processing self.post_init() def allocate_kv_cache(self, batch_size, max_seq_len, inp_seq_len): for layer in self.layers: layer.allocate_kv_cache(batch_size, max_seq_len, inp_seq_len) def reorder_kv_cache(self, beam_idx: torch.LongTensor): return tuple(layer.reorder_kv_cache(beam_idx) for layer in self.layers) def update_sincos_cache(self, seq_len): for layer in self.layers: layer.update_sincos_cache(seq_len) def forward( self, input_ids: Optional[torch.LongTensor] = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None, inputs_embeds: Optional[torch.FloatTensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, cache_position: Optional[torch.LongTensor] = None, token_idx: Optional[torch.Tensor] = None, attn_softmax_bf16: Optional[bool] = False, reuse_cache: Optional[bool] = False, use_flash_attention: Optional[bool] = False, flash_attention_recompute: Optional[bool] = False, flash_attention_causal_mask: Optional[bool] = False, flash_attention_fast_softmax: Optional[bool] = False, valid_sequence_lengths: torch.Tensor = None, cache_idx: int = None, lazy_mode: Optional[bool] = True, num_virtual_tokens: int = None, attn_batch_split: int = 1, **kwargs, ) -> BaseModelOutputWithPast: """ Copied from LlamaModel.forward: https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/modeling_llama.py The only differences are: - add new args token_idx - add new args attn_softmax_bf16 - add new args reuse_cache - add new args use_flash_attention - add new arg flash_attention_recompute - add new arg flash_attention_causal_mask - add new arg flash_attention_fast_softmax - add new arg lazy_mode """ 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 if (input_ids is None) ^ (inputs_embeds is not None): raise ValueError("You must specify exactly one of input_ids or inputs_embeds") elif input_ids is not None: batch_size, seq_length = input_ids.shape[:2] elif inputs_embeds is not None: batch_size, seq_length = inputs_embeds.shape[:2] else: raise ValueError("You have to specify either input_ids or inputs_embeds") if hasattr(self.config, "use_fused_rope") and self.config.use_fused_rope is False: global has_fused_rope has_fused_rope = False if hasattr(self.config, "use_fused_rms_norm") and self.config.use_fused_rms_norm is False: global has_fused_rms_norm has_fused_rms_norm = False if self.gradient_checkpointing and self.training and use_cache: logger.warning_once( "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`." ) use_cache = False if inputs_embeds is None: inputs_embeds = self.embed_tokens(input_ids) ignore_cache_position = True # Ignoring cache position for HPU use_new_cache = False # Ignoring new Cache path for HPU past_seen_tokens = 0 if past_key_values is not None and use_cache: # kept for BC (cache positions) if reuse_cache: if isinstance(past_key_values[0][0], torch.Tensor): past_seen_tokens = past_key_values[0][0].shape[2] else: past_seen_tokens = past_key_values[0][0][2] else: if use_new_cache: if not isinstance(past_key_values, StaticCache): past_key_values = DynamicCache.from_legacy_cache(past_key_values) past_seen_tokens = past_key_values.get_seq_length() else: past_seen_tokens = past_key_values[0][0].shape[2] if ignore_cache_position is False: if cache_position is None: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 cache_position = torch.arange( past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device ) if position_ids is None and cache_position: position_ids = cache_position.unsqueeze(0) else: if position_ids is None: position_ids = torch.arange( past_seen_tokens, seq_length + past_seen_tokens, dtype=torch.long, device=inputs_embeds.device ) position_ids = position_ids.unsqueeze(0) cache_position = None # HPU specific mask generation if ignore_cache_position: causal_mask = _gaudi_prepare_4d_causal_attention_mask( attention_mask, input_ids.shape if input_ids is not None else (batch_size, seq_length), inputs_embeds, past_seen_tokens, ) else: causal_mask = self._update_causal_mask(attention_mask, inputs_embeds, cache_position, past_seen_tokens) # embed positions hidden_states = inputs_embeds # create position embeddings to be shared across the decoder layers position_embeddings = None # self.rotary_emb(hidden_states, position_ids) # decoder layers all_hidden_states = () if output_hidden_states else None all_self_attns = () if output_attentions else None next_decoder_cache = () if not use_new_cache else None if lazy_mode: htcore.mark_step() split_prompt = False prev_layer_residual = None if attn_batch_split > 1 and past_key_values is None: # Calculate split sizes to handle cases where batch size is not divisible by attn_batch_split batch_size = hidden_states.size(0) base_split_size = batch_size // attn_batch_split remainder = batch_size % attn_batch_split split_sizes = [base_split_size + 1 if i < remainder else base_split_size for i in range(attn_batch_split)] # Split tensors using the calculated sizes hidden_states_split = torch.split(hidden_states, split_sizes, dim=0) split_prompt = True for layer_idx, decoder_layer in enumerate(self.layers[: self.config.num_hidden_layers]): if ( lazy_mode and not self.training and (torch.distributed.is_initialized() is False or torch.distributed.get_world_size() == 1) ): htcore.mark_step() if output_hidden_states: all_hidden_states += (hidden_states,) if self.gradient_checkpointing and self.training: layer_outputs = self._gradient_checkpointing_func( partial(decoder_layer.__call__, **kwargs), hidden_states, causal_mask, position_ids, past_key_values, output_attentions, use_cache, cache_position, position_embeddings, None, attn_softmax_bf16, False, use_flash_attention, flash_attention_recompute, flash_attention_causal_mask, flash_attention_fast_softmax, valid_sequence_lengths, None, ) hidden_states = layer_outputs[0] else: # Calling the layer with positional arguments # This is a workaround for an issue with DeepSpeed where # it cannot handle keyword arguments and throws a RuntimError use_prev_layer_residual = attn_batch_split > 1 and past_key_values is None layer_prev_layer_residual = prev_layer_residual if use_prev_layer_residual else None layer_hidden_states = hidden_states_split if split_prompt else hidden_states past_key_value = None if past_key_values is None else past_key_values[layer_idx] layer_outputs = decoder_layer( layer_hidden_states, causal_mask, position_ids, past_key_value, output_attentions, use_cache, cache_position, position_embeddings, token_idx, attn_softmax_bf16, reuse_cache, use_flash_attention, flash_attention_recompute, flash_attention_causal_mask, flash_attention_fast_softmax, valid_sequence_lengths, cache_idx, num_virtual_tokens, attn_batch_split, layer_prev_layer_residual, ) if use_prev_layer_residual: index = 1 + int(use_cache) + int(output_attentions) prev_layer_residual = layer_outputs[index] if split_prompt: hidden_states_split = layer_outputs[0] else: 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],) hidden_states = self.norm(hidden_states) # add hidden states from the last decoder layer if output_hidden_states: all_hidden_states += (hidden_states,) next_cache = next_decoder_cache if use_cache else None if not use_new_cache and isinstance(next_cache, Cache): next_cache = next_cache.to_legacy_cache() return BaseModelOutputWithPast( last_hidden_state=hidden_states, past_key_values=next_cache, hidden_states=all_hidden_states, attentions=all_self_attns, ) class GaudiLlamaForCausalLM(LlamaForCausalLM): """ Inherits from LlamaForCausalLM: https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/modeling_llama.py The only differences are: - add new args token_idx - add token_idx into model_inputs - from step2 when enable KV cache, slice next_input_ids from input_ids base on the token_idx - from step2 when enable KV cache, slice next_position_ids from position_ids base on the token_idx - add new args attn_softmax_bf16 - add new args reuse_cache """ def __init__(self, config, parallel_strategy: DistributedStrategy = NoOpStrategy): config.parallel_strategy = parallel_strategy super().__init__(config) if parallel_state.sequence_parallel_is_initialized() and parallel_state.get_sequence_parallel_world_size() > 1: from ....distributed.contextparallel import ForCausalLMContextParallelLoss self._loss_function = ForCausalLMContextParallelLoss def allocate_kv_cache(self, batch_size, max_seq_len, inp_seq_len): self.model.allocate_kv_cache(batch_size, max_seq_len, inp_seq_len) def reorder_kv_cache(self, beam_idx: torch.LongTensor): return self.model.reorder_kv_cache(beam_idx) def update_sincos_cache(self, seq_len): self.model.update_sincos_cache(seq_len) def forward( self, input_ids: Optional[torch.LongTensor] = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_values: Optional[Union[Cache, 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, cache_position: Optional[torch.LongTensor] = None, logits_to_keep: Union[int, torch.Tensor] = 0, token_idx: Optional[torch.Tensor] = None, trim_logits: Optional[bool] = False, attn_softmax_bf16: Optional[bool] = False, reuse_cache: Optional[bool] = False, use_flash_attention: Optional[bool] = False, flash_attention_recompute: Optional[bool] = False, flash_attention_causal_mask: Optional[bool] = False, flash_attention_fast_softmax: Optional[bool] = False, valid_sequence_lengths: torch.Tensor = None, cache_idx: int = None, lazy_mode: Optional[bool] = True, num_virtual_tokens: int = None, attn_batch_split: int = 1, **kwargs: Unpack[KwargsForCausalLM], ) -> CausalLMOutputWithPast: 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 ) if self.generation_config.use_fused_rope is False: global has_fused_rope has_fused_rope = False # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn) outputs: BaseModelOutputWithPast = 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, cache_position=cache_position, token_idx=token_idx, attn_softmax_bf16=attn_softmax_bf16, reuse_cache=reuse_cache, use_flash_attention=use_flash_attention, flash_attention_recompute=flash_attention_recompute, flash_attention_causal_mask=flash_attention_causal_mask, flash_attention_fast_softmax=flash_attention_fast_softmax, valid_sequence_lengths=valid_sequence_lengths, cache_idx=cache_idx, lazy_mode=lazy_mode, num_virtual_tokens=num_virtual_tokens, attn_batch_split=attn_batch_split, **kwargs, ) hidden_states = outputs.last_hidden_state _, seq_len, _ = hidden_states.shape if seq_len > 1 and trim_logits and not self.training: if token_idx is not None: hidden_states = hidden_states.index_select(1, token_idx - 1) else: hidden_states = hidden_states[:, -1, :] # Only compute necessary logits, and do not upcast them to float if we are not computing the loss slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep logits = self.lm_head(hidden_states[:, slice_indices, :]).float() loss = None if labels is not None: loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs) return CausalLMOutputWithPast( loss=loss, logits=logits, past_key_values=outputs.past_key_values, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) @staticmethod def _reorder_cache( past: Tuple[Tuple[torch.Tensor, torch.Tensor], ...], beam_idx: torch.LongTensor ) -> Tuple[Tuple[torch.Tensor, torch.Tensor], ...]: """ This function is used to re-order the `past_key_values` cache if [`~PreTrainedModel.beam_search`] or [`~PreTrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct beam_idx at every generation step. Output shares the same memory storage as `past`. """ return tuple( ( layer_past[0].index_select(0, beam_idx.to(layer_past[0].device)), layer_past[1].index_select(0, beam_idx.to(layer_past[1].device)), ) for layer_past in past ) def prepare_inputs_for_generation( self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, cache_position=None, position_ids=None, use_cache=True, num_logits_to_keep=None, token_idx=None, **kwargs, ): reuse_cache = kwargs.get("reuse_cache") bucket_internal = kwargs.get("bucket_internal") if past_key_values is not None: if token_idx is not None: idx = token_idx + kwargs.get("inputs_embeds_offset", 0) - 1 input_ids = torch.index_select(input_ids, 1, idx) else: if inputs_embeds is not None: # Exception 1 input_ids = input_ids[:, -cache_position.shape[0] :] elif ( input_ids.shape[1] != cache_position.shape[0] ): # Default case (the "else", a no op, is Exception 2) input_ids = input_ids[:, cache_position] elif (reuse_cache or bucket_internal) and token_idx is not None: # KV cache is pre allocated with reuse cache or will be padded with bucket internal # hence for the 1st token we can slice the inputs till token idx for the fwd pass. input_ids = input_ids[:, :token_idx] attention_mask = attention_mask[:, :token_idx] 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) if past_key_values: if token_idx is not None: position_ids = torch.index_select(position_ids, 1, token_idx - 1) else: position_ids = position_ids[:, -input_ids.shape[1] :] # This `clone` call is needed to avoid recapturing cuda graphs with `torch.compile`'s `mode="reduce-overhead`, as otherwise the input `position_ids` would have various stride during the decoding. Here, simply using `.contiguous()` is not sufficient as in the batch size = 1 case, `position_ids` is already contiguous but with varying stride which retriggers a capture. position_ids = position_ids.clone(memory_format=torch.contiguous_format) # keep cache_position implementation as None for HPU cache_position = None # if `inputs_embeds` are passed, we only want to use them in the 1st generation step if inputs_embeds is not None and past_key_values is None: model_inputs = {"inputs_embeds": inputs_embeds} else: model_inputs = {"input_ids": input_ids.clone(memory_format=torch.contiguous_format)} if num_logits_to_keep is not None: model_inputs["num_logits_to_keep"] = num_logits_to_keep if bucket_internal and reuse_cache is not True: # update input with kv cache len to capture padding changes during internal bucketing without cache reuse model_inputs["kv_cache_len"] = kwargs.get("kv_cache_len") model_inputs.update( { "position_ids": position_ids, "cache_position": cache_position, "past_key_values": past_key_values, "use_cache": use_cache, "attention_mask": attention_mask, "token_idx": token_idx, "trim_logits": kwargs.get("trim_logits"), "attn_softmax_bf16": kwargs.get("attn_softmax_bf16"), "reuse_cache": reuse_cache, "use_flash_attention": kwargs.get("use_flash_attention"), "flash_attention_recompute": kwargs.get("flash_attention_recompute"), "flash_attention_causal_mask": kwargs.get("flash_attention_causal_mask"), "flash_attention_fast_softmax": kwargs.get("flash_attention_fast_softmax"), "valid_sequence_lengths": kwargs.get("valid_sequence_lengths"), "cache_idx": kwargs.get("cache_idx"), "lazy_mode": kwargs.get("lazy_mode"), "num_virtual_tokens": kwargs.get("num_virtual_tokens"), "attn_batch_split": kwargs.get("attn_batch_split"), } ) return model_inputs # Transformer4.43 use new Cache mechanism while Gaudi is not. # Adding _reorder_cache back to support HPU. @staticmethod def _reorder_cache(past_key_values, beam_idx): reordered_past = () for layer_past in past_key_values: reordered_past += ( tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past), ) return reordered_past def apply_customized_rope(q, k, cos, sin, position_ids, training=True): if q.device.type == "hpu" and has_fused_rope: return apply_customized_rope_module(q, k, cos, sin, position_ids, training) else: return apply_rotary_pos_emb(q, k, cos[position_ids], sin[position_ids])