From 5ff2dc9176d3473f3febdc5ad1813f4b1f54ec0a Mon Sep 17 00:00:00 2001 From: OlivierDehaene <23298448+OlivierDehaene@users.noreply.github.com> Date: Thu, 1 Jun 2023 10:05:24 +0200 Subject: [PATCH] wip --- .../custom_modeling/flash_llama_modeling.py | 6 +- .../custom_modeling/flash_neox_modeling.py | 6 +- .../custom_modeling/flash_rw_modeling.py | 272 +++++++++------- .../flash_santacoder_modeling.py | 6 +- .../models/flash_causal_lm.py | 297 +++++++++--------- 5 files changed, 310 insertions(+), 277 deletions(-) diff --git a/server/text_generation_server/models/custom_modeling/flash_llama_modeling.py b/server/text_generation_server/models/custom_modeling/flash_llama_modeling.py index 8a35ffa8..de1627b1 100644 --- a/server/text_generation_server/models/custom_modeling/flash_llama_modeling.py +++ b/server/text_generation_server/models/custom_modeling/flash_llama_modeling.py @@ -26,7 +26,7 @@ from transformers.activations import ACT2FN from typing import Optional # Flash attention imports -import flash_attn_cuda +import flash_attn_cuda_modif import dropout_layer_norm from text_generation_server.utils.layers import ( @@ -149,7 +149,7 @@ class FlashLlamaAttention(torch.nn.Module): # output attn_output = torch.empty_like(qkv[:, 0]) # flash attention - flash_attn_cuda.fwd( + flash_attn_cuda_modif.fwd( qkv[:, 0], qkv[:, 1], qkv[:, 2], @@ -175,7 +175,7 @@ class FlashLlamaAttention(torch.nn.Module): # output attn_output = torch.empty_like(query) # flash attention - flash_attn_cuda.fwd( + flash_attn_cuda_modif.fwd( query, layer_past[:, 0], layer_past[:, 1], diff --git a/server/text_generation_server/models/custom_modeling/flash_neox_modeling.py b/server/text_generation_server/models/custom_modeling/flash_neox_modeling.py index 0fe43bcb..e0aa2cb8 100644 --- a/server/text_generation_server/models/custom_modeling/flash_neox_modeling.py +++ b/server/text_generation_server/models/custom_modeling/flash_neox_modeling.py @@ -28,7 +28,7 @@ from transformers.models.gpt_neox import GPTNeoXConfig from typing import Optional # Flash attention imports -import flash_attn_cuda +import flash_attn_cuda_modif from text_generation_server.utils.layers import ( TensorParallelRowLinear, @@ -134,7 +134,7 @@ class FlashNeoxAttention(torch.nn.Module): # output attn_output = torch.empty_like(qkv[:, 0]) # flash attention - flash_attn_cuda.fwd( + flash_attn_cuda_modif.fwd( qkv[:, 0], qkv[:, 1], qkv[:, 2], @@ -160,7 +160,7 @@ class FlashNeoxAttention(torch.nn.Module): # output attn_output = torch.empty_like(query) # flash attention - flash_attn_cuda.fwd( + flash_attn_cuda_modif.fwd( query, layer_past[:, 0], layer_past[:, 1], diff --git a/server/text_generation_server/models/custom_modeling/flash_rw_modeling.py b/server/text_generation_server/models/custom_modeling/flash_rw_modeling.py index 55195162..f727f8ad 100644 --- a/server/text_generation_server/models/custom_modeling/flash_rw_modeling.py +++ b/server/text_generation_server/models/custom_modeling/flash_rw_modeling.py @@ -1,13 +1,14 @@ import torch import torch.distributed +from loguru import logger from torch import nn from transformers.modeling_utils import PreTrainedModel from transformers.configuration_utils import PretrainedConfig from typing import Optional # Flash attention imports -import flash_attn_cuda +import flash_attn_cuda_modif from text_generation_server.utils.layers import ( TensorParallelRowLinear, @@ -42,25 +43,25 @@ class RWConfig(PretrainedConfig): } def __init__( - self, - model_type="RefinedWeb", - vocab_size=250880, - hidden_size=64, - n_layer=2, - n_head=8, - layer_norm_epsilon=1e-5, - initializer_range=0.02, - use_cache=True, - bos_token_id=1, - eos_token_id=2, - hidden_dropout=0.0, - attention_dropout=0.0, - n_head_kv=None, - multi_query=False, - alibi=False, - bias=False, - parallel_attn=False, - **kwargs, + self, + model_type="RefinedWeb", + vocab_size=250880, + hidden_size=64, + n_layer=2, + n_head=8, + layer_norm_epsilon=1e-5, + initializer_range=0.02, + use_cache=True, + bos_token_id=1, + eos_token_id=2, + hidden_dropout=0.0, + attention_dropout=0.0, + n_head_kv=None, + multi_query=False, + alibi=False, + bias=False, + parallel_attn=False, + **kwargs, ): if alibi: raise NotImplementedError( @@ -126,15 +127,19 @@ class FlashRWAttention(torch.nn.Module): ) def forward( - self, - hidden_states, - cos, - sin, - cu_seqlens, - max_s, - layer_past, - layer_past_present_indices, - cu_seqlens_q, + self, + hidden_states, + cos, + sin, + start_seq, + end_seq, + start_seq_q, + end_seq_q, + max_s, + layer_past, + layer_past_present_indices, + prefill, + past_stream ): qkv = self.query_key_value(hidden_states) @@ -153,22 +158,26 @@ class FlashRWAttention(torch.nn.Module): self.rotary_emb(kv[:, 0], cos, sin) # Prefill - if layer_past_present_indices is None: - # Copy to layer past - layer_past[...] = kv + if prefill: + past_stream.wait_stream(torch.cuda.current_stream()) + with torch.cuda.stream(past_stream): + # Copy to layer past + layer_past[layer_past_present_indices] = kv # Expand to query shape kv = kv.expand(-1, 2, self.num_heads, self.head_size) # output attn_output = torch.empty_like(query) # flash attention - flash_attn_cuda.fwd( + flash_attn_cuda_modif.fwd( query, kv[:, 0], kv[:, 1], attn_output, - cu_seqlens, - cu_seqlens, + start_seq, + end_seq, + start_seq, + end_seq, max_s, max_s, 0.0, @@ -181,6 +190,7 @@ class FlashRWAttention(torch.nn.Module): ) # Decode else: + torch.cuda.current_stream().wait_stream(past_stream) # Add present to the layer_past tensor at the correct indices layer_past[layer_past_present_indices] = kv # Expand to query shape @@ -189,13 +199,15 @@ class FlashRWAttention(torch.nn.Module): # output attn_output = torch.empty_like(query) # flash attention - flash_attn_cuda.fwd( + flash_attn_cuda_modif.fwd( query, kv[:, 0], kv[:, 1], attn_output, - cu_seqlens_q, - cu_seqlens, + start_seq_q, + end_seq_q, + start_seq, + end_seq, 1, max_s, 0.0, @@ -257,15 +269,15 @@ class FlashRWLargeAttention(torch.nn.Module): ) def forward( - self, - hidden_states, - cos, - sin, - cu_seqlens, - max_s, - layer_past, - layer_past_present_indices, - cu_seqlens_q, + self, + hidden_states, + cos, + sin, + cu_seqlens, + max_s, + layer_past, + layer_past_present_indices, + cu_seqlens_q, ): qkv = self.query_key_value(hidden_states) qkv = qkv.view(-1, self.num_groups, self.num_heads + 2, self.head_size) @@ -296,7 +308,7 @@ class FlashRWLargeAttention(torch.nn.Module): # output attn_output = torch.empty_like(query) # flash attention - flash_attn_cuda.fwd( + flash_attn_cuda_modif.fwd( query, kv[:, :, 0], kv[:, :, 1], @@ -327,7 +339,7 @@ class FlashRWLargeAttention(torch.nn.Module): # output attn_output = torch.empty_like(query) # flash attention - flash_attn_cuda.fwd( + flash_attn_cuda_modif.fwd( query, kv[:, :, 0], kv[:, :, 1], @@ -412,16 +424,20 @@ class FlashRWLayer(nn.Module): self.process_group = weights.process_group def forward( - self, - hidden_states, - residual, - cos, - sin, - cu_seqlens, - max_s, - layer_past, - layer_past_present_indices, - cu_seqlens_q, + self, + hidden_states, + residual, + cos, + sin, + start_seq, + end_seq, + start_seq_q, + end_seq_q, + max_s, + layer_past, + layer_past_present_indices, + prefill, + past_stream, ): if self.parallel_attn: ln_hidden_states, residual = self.input_layernorm(hidden_states, residual) @@ -430,11 +446,15 @@ class FlashRWLayer(nn.Module): ln_hidden_states, cos, sin, - cu_seqlens, + start_seq, + end_seq, + start_seq_q, + end_seq_q, max_s, layer_past, layer_past_present_indices, - cu_seqlens_q, + prefill, + past_stream ) mlp_output = self.mlp(ln_hidden_states) @@ -450,11 +470,14 @@ class FlashRWLayer(nn.Module): hidden_states, cos, sin, - cu_seqlens, + start_seq, + end_seq, + start_seq_q, + end_seq_q, max_s, layer_past, layer_past_present_indices, - cu_seqlens_q, + prefill, ) hidden_states, residual = self.post_attention_layernorm( @@ -493,16 +516,16 @@ class FlashRWLargeLayer(nn.Module): self.process_group = weights.process_group def forward( - self, - hidden_states, - residual, - cos, - sin, - cu_seqlens, - max_s, - layer_past, - layer_past_present_indices, - cu_seqlens_q, + self, + hidden_states, + residual, + cos, + sin, + cu_seqlens, + max_s, + layer_past, + layer_past_present_indices, + cu_seqlens_q, ): ln_attn, residual = self.ln_attn(hidden_states, residual) ln_mlp, _ = self.ln_mlp(residual) @@ -554,6 +577,7 @@ class FlashRWModel(FlashRWPreTrainedModel): self.h[0].self_attention.head_size, ) elif config.model_type == "RefinedWeb": + raise NotImplementedError self.h = nn.ModuleList( [ FlashRWLargeLayer(layer_id, config, weights) @@ -577,38 +601,55 @@ class FlashRWModel(FlashRWPreTrainedModel): ) self.head_size = self.h[0].self_attention.head_size + self.past_stream = torch.cuda.Stream() def forward( - self, - input_ids, - position_ids, - cu_seqlens, - cu_seqlens_q, - max_s, - past_key_values=None, - pre_allocate_past_size: Optional[int] = None, + self, + input_ids, + position_ids, + start_seq, + end_seq, + start_seq_q, + end_seq_q, + max_s, + past_key_values=None, + pre_allocate_past_size: Optional[int] = None, ): hidden_states = self.word_embeddings(input_ids) # Prefill if past_key_values is None: - # Create past tensor - past_key_values = hidden_states.new_empty( - ( - len(self.h), - len(hidden_states) - if pre_allocate_past_size is None - else pre_allocate_past_size, - *self.cache_size, + assert pre_allocate_past_size is not None + + prefill = True + + with torch.cuda.stream(self.past_stream): + # Create past tensor + past_key_values = hidden_states.new_zeros( + ( + len(self.h), + pre_allocate_past_size, + *self.cache_size, + ) ) - ) - layer_past_present_indices = None - slice_past_index = len(hidden_states) + seq_indices = [] + for s, e in zip(start_seq, end_seq): + seq_indices.append( + torch.arange( + s, + e, + dtype=torch.int64, + device=self.device + ) + ) + layer_past_present_indices = torch.cat(seq_indices) + from loguru import logger + logger.error(f"layer past: {layer_past_present_indices}") # Decode else: + prefill = False # Create indices from cumulative sequence lengths - layer_past_present_indices = cu_seqlens[1:] - 1 - slice_past_index = None + layer_past_present_indices = end_seq - 1 # Get rotary cos and sin for this forward # Avoid to index in each layer @@ -618,23 +659,20 @@ class FlashRWModel(FlashRWPreTrainedModel): residual = None for i, layer in enumerate(self.h): - # We added padding that we now need to slice - layer_past_key_values = ( - past_key_values[i] - if slice_past_index is None - else past_key_values[i, :slice_past_index] - ) - hidden_states, residual = layer( hidden_states, residual, cos, sin, - cu_seqlens, + start_seq, + end_seq, + start_seq_q, + end_seq_q, max_s, - layer_past_key_values, + past_key_values[i], layer_past_present_indices, - cu_seqlens_q, + prefill, + self.past_stream ) hidden_states, _ = self.ln_f(hidden_states, residual) @@ -653,21 +691,25 @@ class FlashRWForCausalLM(FlashRWPreTrainedModel): ) def forward( - self, - input_ids, - position_ids, - cu_seqlens, - cu_seqlens_q, - max_s, - past_key_values: Optional[torch.Tensor] = None, - pre_allocate_past_size: Optional[int] = None, - lm_head_indices: Optional[torch.Tensor] = None, + self, + input_ids, + position_ids, + start_seq, + end_seq, + start_seq_q, + end_seq_q, + max_s, + past_key_values: Optional[torch.Tensor] = None, + pre_allocate_past_size: Optional[int] = None, + lm_head_indices: Optional[torch.Tensor] = None, ): hidden_states, present = self.transformer( input_ids, position_ids, - cu_seqlens, - cu_seqlens_q, + start_seq, + end_seq, + start_seq_q, + end_seq_q, max_s, past_key_values, pre_allocate_past_size, diff --git a/server/text_generation_server/models/custom_modeling/flash_santacoder_modeling.py b/server/text_generation_server/models/custom_modeling/flash_santacoder_modeling.py index 888a6066..c22aac65 100644 --- a/server/text_generation_server/models/custom_modeling/flash_santacoder_modeling.py +++ b/server/text_generation_server/models/custom_modeling/flash_santacoder_modeling.py @@ -6,7 +6,7 @@ from transformers.activations import ACT2FN from typing import Optional # Flash attention imports -import flash_attn_cuda +import flash_attn_cuda_modif from text_generation_server.utils.layers import ( TensorParallelRowLinear, TensorParallelColumnLinear, @@ -175,7 +175,7 @@ class FlashMQAttention(torch.nn.Module): # output attn_output = torch.empty_like(query) # flash attention - flash_attn_cuda.fwd( + flash_attn_cuda_modif.fwd( query, key_value[:, 0], key_value[:, 1], @@ -202,7 +202,7 @@ class FlashMQAttention(torch.nn.Module): # output attn_output = torch.empty_like(query) # flash attention - flash_attn_cuda.fwd( + flash_attn_cuda_modif.fwd( query, key_value[:, 0], key_value[:, 1], diff --git a/server/text_generation_server/models/flash_causal_lm.py b/server/text_generation_server/models/flash_causal_lm.py index a2ad2d5e..d3323cb8 100644 --- a/server/text_generation_server/models/flash_causal_lm.py +++ b/server/text_generation_server/models/flash_causal_lm.py @@ -34,10 +34,18 @@ class FlashCausalLMBatch(Batch): input_ids: torch.Tensor position_ids: torch.Tensor - # cumulative sequence lengths - cu_seqlens: torch.Tensor - # cumulative query sequence lengths, only used in decode - cu_seqlens_q: Optional[torch.Tensor] + # tensor of length b holding starting offset of each sequence + start_seq: torch.Tensor + # tensor of length b holding ending offset of each sequence + end_seq: torch.Tensor + # tensor of length b holding starting offset of each sequence, only used in prefill + start_seq_prefill: Optional[torch.Tensor] + # tensor of length b holding ending offset of each sequence, only used in prefill + end_seq_prefill: Optional[torch.Tensor] + # tensor of length b holding starting offset of each query sequence, only used in decode + start_seq_q: Optional[torch.Tensor] + # tensor of length b holding ending offset of each query sequence, only used in decode + end_seq_q: Optional[torch.Tensor] # past key values, only used in decode past_key_values: Optional[torch.Tensor] max_seqlen: int @@ -73,11 +81,11 @@ class FlashCausalLMBatch(Batch): @classmethod def from_pb( - cls, - pb: generate_pb2.Batch, - tokenizer: PreTrainedTokenizerBase, - dtype: torch.dtype, - device: torch.device, + cls, + pb: generate_pb2.Batch, + tokenizer: PreTrainedTokenizerBase, + dtype: torch.dtype, + device: torch.device, ) -> "FlashCausalLMBatch": batch_inputs = [] max_truncation = 0 @@ -90,7 +98,10 @@ class FlashCausalLMBatch(Batch): )["input_ids"] position_ids = [] - cu_seqlens = [0] + start_seq = [] + end_seq = [] + start_seq_prefill = [] + end_seq_prefill = [] max_seqlen = 0 input_lengths = [] @@ -110,9 +121,9 @@ class FlashCausalLMBatch(Batch): # Cumulative length cumulative_length = 0 + cumulative_max_length = 0 prefill_out_cumulative_length = 0 - max_tokens = 0 max_length = 0 # Parse batch @@ -138,7 +149,10 @@ class FlashCausalLMBatch(Batch): position_ids.append(request_position_ids) # Add cumulative lengths of all previous inputs - cu_seqlens.append(cumulative_length + input_length) + start_seq_prefill.append(cumulative_length) + end_seq_prefill.append(cumulative_length + input_length) + start_seq.append(cumulative_max_length) + end_seq.append(cumulative_max_length + input_length) next_token_chooser_parameters.append(r.parameters) @@ -169,8 +183,9 @@ class FlashCausalLMBatch(Batch): prefill_out_cumulative_length += 1 # Update + # Remove one as the first token des not have a past cumulative_length += input_length - max_tokens += input_length + max_new_tokens + cumulative_max_length += input_length + max_new_tokens - 1 max_length = max(max_length, input_length + max_new_tokens) next_token_chooser = HeterogeneousNextTokenChooser.from_pb( @@ -197,13 +212,20 @@ class FlashCausalLMBatch(Batch): all_input_ids_tensor, dtype=torch.int64, device=device ) position_ids = torch.tensor(position_ids, dtype=torch.int32, device=device) - cu_seqlens = torch.tensor(cu_seqlens, device=device, dtype=torch.int32) + start_seq = torch.tensor(start_seq, device=device, dtype=torch.int32) + end_seq = torch.tensor(end_seq, device=device, dtype=torch.int32) + if len(pb.requests) > 1: + start_seq_prefill = torch.tensor(start_seq_prefill, device=device, dtype=torch.int32) + end_seq_prefill = torch.tensor(end_seq_prefill, device=device, dtype=torch.int32) + else: + start_seq_prefill = start_seq + end_seq_prefill = end_seq if all_prefill_logprobs: prefill_head_indices = None - prefill_next_token_indices = cu_seqlens[1:] - 1 + prefill_next_token_indices = end_seq - 1 elif no_prefill_logprobs: - prefill_head_indices = cu_seqlens[1:] - 1 + prefill_head_indices = end_seq - 1 prefill_next_token_indices = None else: prefill_head_indices = torch.tensor( @@ -219,8 +241,12 @@ class FlashCausalLMBatch(Batch): requests_idx_mapping=requests_idx_mapping, input_ids=input_ids, position_ids=position_ids, - cu_seqlens=cu_seqlens, - cu_seqlens_q=None, + start_seq=start_seq, + end_seq=end_seq, + start_seq_prefill=start_seq_prefill, + end_seq_prefill=end_seq_prefill, + start_seq_q=None, + end_seq_q=None, max_seqlen=max_seqlen, prefill_head_indices=prefill_head_indices, prefill_next_token_indices=prefill_next_token_indices, @@ -233,7 +259,7 @@ class FlashCausalLMBatch(Batch): all_input_ids_tensor=all_input_ids_tensor, next_token_chooser=next_token_chooser, stopping_criterias=stopping_criterias, - max_tokens=max_tokens, + max_tokens=cumulative_max_length, ) @tracer.start_as_current_span("filter") @@ -247,7 +273,7 @@ class FlashCausalLMBatch(Batch): single_request = len(request_ids) == 1 # Cumulative length - cumulative_length = 0 + cumulative_max_length = 0 # New values after filtering requests_idx_mapping = {} @@ -256,8 +282,10 @@ class FlashCausalLMBatch(Batch): indices = [] # Create on CPU to only move to GPU once instead of at every copy - cu_seqlens = torch.zeros(len(request_ids) + 1, dtype=torch.int32) - cu_seqlens_q = self.cu_seqlens_q[: len(request_ids) + 1] + start_seq = torch.empty(len(request_ids), dtype=torch.int32) + end_seq = torch.empty(len(request_ids), dtype=torch.int32) + start_seq_q = self.start_seq_q[: len(request_ids)] + end_seq_q = self.end_seq_q[: len(request_ids)] max_seqlen = 0 past_key_values = [] @@ -270,8 +298,6 @@ class FlashCausalLMBatch(Batch): stopping_criterias = [] - max_tokens = 0 - for i, request_id in enumerate(request_ids): idx = self.requests_idx_mapping[request_id] indices.append(idx) @@ -281,16 +307,8 @@ class FlashCausalLMBatch(Batch): # Get length request_input_length = self.input_lengths[idx] - - # Copy to tensor (CPU) - cu_seqlens[i + 1] = cumulative_length + request_input_length max_seqlen = max(max_seqlen, request_input_length) - # Slice from past - past_key_values.append( - self.past_key_values[:, self.cu_seqlens[idx] : self.cu_seqlens[idx + 1]] - ) - all_input_ids.append(self.all_input_ids[idx]) input_lengths.append(request_input_length) @@ -300,30 +318,19 @@ class FlashCausalLMBatch(Batch): stopping_criteria = self.stopping_criterias[idx] stopping_criterias.append(stopping_criteria) - cumulative_length += request_input_length - max_tokens += request_input_length + ( - stopping_criteria.max_new_tokens - stopping_criteria.current_tokens + remaining_tokens = stopping_criteria.max_new_tokens - stopping_criteria.current_tokens + + # Copy to tensor (CPU) + start_seq[i] = cumulative_max_length + end_seq[i] = cumulative_max_length + request_input_length + + # Slice from past + past_key_values.append( + self.past_key_values[:, + self.start_seq[idx]: self.end_seq[idx] + remaining_tokens - 1] ) - if single_request: - # Preallocate tensor for bs = 1 case - past_key_values = F.pad( - past_key_values[0], - ( - 0, - 0, - 0, - 0, - 0, - 0, - 0, - stopping_criterias[0].max_new_tokens - - stopping_criterias[0].current_tokens, - ), - ) - else: - # Cat all past - past_key_values = torch.cat(past_key_values, dim=1) + cumulative_max_length += request_input_length + remaining_tokens - 1 # Index into tensors input_ids = self.input_ids[indices] @@ -331,8 +338,15 @@ class FlashCausalLMBatch(Batch): all_input_ids_tensor = self.all_input_ids_tensor[indices] next_token_chooser = self.next_token_chooser.filter(indices) + if single_request: + past_key_values = past_key_values[0] + else: + # Cat all past + past_key_values = torch.cat(past_key_values, dim=1) + # Move to GPU now that we have the whole tensor - cu_seqlens = cu_seqlens.to(self.cu_seqlens.device) + start_seq = start_seq.to(self.start_seq.device) + end_seq = end_seq.to(self.start_seq.device) return FlashCausalLMBatch( batch_id=self.batch_id, @@ -340,8 +354,12 @@ class FlashCausalLMBatch(Batch): requests_idx_mapping=requests_idx_mapping, input_ids=input_ids, position_ids=position_ids, - cu_seqlens=cu_seqlens, - cu_seqlens_q=cu_seqlens_q, + start_seq=start_seq, + end_seq=end_seq, + start_seq_prefill=None, + end_seq_prefill=None, + start_seq_q=start_seq_q, + end_seq_q=end_seq_q, max_seqlen=max_seqlen, prefill_head_indices=None, prefill_next_token_indices=None, @@ -354,7 +372,7 @@ class FlashCausalLMBatch(Batch): all_input_ids_tensor=all_input_ids_tensor, next_token_chooser=next_token_chooser, stopping_criterias=stopping_criterias, - max_tokens=max_tokens, + max_tokens=cumulative_max_length, ) @classmethod @@ -365,18 +383,25 @@ class FlashCausalLMBatch(Batch): requests_idx_mapping = {} total_batch_size = sum([len(b) for b in batches]) + total_tokens = sum(b.max_tokens for b in batches) dtype = batches[0].past_key_values.dtype device = batches[0].input_ids.device input_ids = batches[0].input_ids.new_empty(total_batch_size) position_ids = batches[0].position_ids.new_empty(total_batch_size) - cu_seqlens = [0] - cu_seqlens_q = torch.arange( - 0, total_batch_size + 1, device=device, dtype=torch.int32 + start_seq = batches[0].start_seq.new_empty(total_batch_size) + end_seq = batches[0].end_seq.new_empty(total_batch_size) + start_seq_q = torch.arange( + 0, total_batch_size, device=device, dtype=torch.int32 ) + end_seq_q = start_seq_q + 1 max_seqlen = 0 - past_key_values = [] + past_key_values = batches[0].past_key_values.new_empty(( + batches[0].past_key_values.shape[0], + total_tokens, + *batches[0].past_key_values.shape[2:] + )) all_input_ids = [] @@ -389,7 +414,6 @@ class FlashCausalLMBatch(Batch): # Cumulative length cumulative_batch_size = 0 - cumulative_length = 0 max_tokens = 0 max_length = 0 @@ -410,18 +434,15 @@ class FlashCausalLMBatch(Batch): input_ids[start_index:end_index] = batch.input_ids position_ids[start_index:end_index] = batch.position_ids - # Add cumulative lengths of all previous inputs - cu_seqlens.extend([l + cumulative_length for l in batch.cu_seqlens[1:]]) - max_seqlen = max(max_seqlen, batch.max_seqlen) + start_seq[start_index:end_index] = batch.start_seq + max_tokens + end_seq[start_index:end_index] = batch.end_seq + max_tokens - if len(batch) != 1: - past_key_values.append(batch.past_key_values) - else: - # past was pre-allocated for this batch - # We need to slice to remove the padding - past_key_values.append( - batch.past_key_values[:, : batch.input_lengths[0]] - ) + past_key_values[ + :, + max_tokens: max_tokens + batch.max_tokens + ] = batch.past_key_values + + max_seqlen = max(max_seqlen, batch.max_seqlen) all_input_ids.extend(batch.all_input_ids) @@ -433,7 +454,6 @@ class FlashCausalLMBatch(Batch): stopping_criterias.extend(batch.stopping_criterias) # Update - cumulative_length += batch.cu_seqlens[-1] cumulative_batch_size += len(batch) max_tokens += batch.max_tokens max_length = max( @@ -458,16 +478,11 @@ class FlashCausalLMBatch(Batch): end_index = cumulative_batch_size + len(batch) all_input_ids_tensor[ - start_index:end_index, : batch.all_input_ids_tensor.shape[1] + start_index:end_index, : batch.all_input_ids_tensor.shape[1] ] = batch.all_input_ids_tensor[:, :max_length] cumulative_batch_size += len(batch) - # Cat past - past_key_values = torch.cat(past_key_values, dim=1) - # Create final tensor on GPU - cu_seqlens = torch.tensor(cu_seqlens, dtype=torch.int32, device=device) - next_token_chooser = HeterogeneousNextTokenChooser.from_pb( next_token_chooser_parameters, dtype=dtype, device=device ) @@ -478,8 +493,12 @@ class FlashCausalLMBatch(Batch): requests_idx_mapping=requests_idx_mapping, input_ids=input_ids, position_ids=position_ids, - cu_seqlens=cu_seqlens, - cu_seqlens_q=cu_seqlens_q, + start_seq=start_seq, + end_seq=end_seq, + start_seq_prefill=None, + end_seq_prefill=None, + start_seq_q=start_seq_q, + end_seq_q=end_seq_q, max_seqlen=max_seqlen, prefill_head_indices=None, prefill_next_token_indices=None, @@ -501,12 +520,12 @@ class FlashCausalLMBatch(Batch): class FlashCausalLM(Model): def __init__( - self, - model_cls: Type[PreTrainedModel], - model_id: str, - revision: Optional[str] = None, - quantize: Optional[str] = None, - trust_remote_code: bool = False, + self, + model_cls: Type[PreTrainedModel], + model_id: str, + revision: Optional[str] = None, + quantize: Optional[str] = None, + trust_remote_code: bool = False, ): if torch.cuda.is_available(): device = torch.device("cuda") @@ -547,22 +566,26 @@ class FlashCausalLM(Model): ) def forward( - self, - input_ids: torch.Tensor, - position_ids: torch.Tensor, - cu_seqlens: torch.Tensor, - cu_seqlens_q: Optional[torch.Tensor], - max_s: int, - past_key_values: Optional = None, - pre_allocate_past_size: Optional[int] = None, - lm_head_indices: Optional[torch.Tensor] = None, + self, + input_ids: torch.Tensor, + position_ids: torch.Tensor, + start_seq: torch.Tensor, + end_seq: torch.Tensor, + start_seq_q: Optional[torch.Tensor], + end_seq_q: Optional[torch.Tensor], + max_s: int, + past_key_values: Optional = None, + pre_allocate_past_size: Optional[int] = None, + lm_head_indices: Optional[torch.Tensor] = None, ) -> Tuple[torch.Tensor, torch.Tensor]: # Model Forward return self.model.forward( input_ids=input_ids, position_ids=position_ids, - cu_seqlens=cu_seqlens, - cu_seqlens_q=cu_seqlens_q, + start_seq=start_seq, + end_seq=end_seq, + start_seq_q=start_seq_q, + end_seq_q=end_seq_q, max_s=max_s, past_key_values=past_key_values, pre_allocate_past_size=pre_allocate_past_size, @@ -571,7 +594,7 @@ class FlashCausalLM(Model): @tracer.start_as_current_span("generate_token") def generate_token( - self, batch: FlashCausalLMBatch + self, batch: FlashCausalLMBatch ) -> Tuple[List[Generation], Optional[FlashCausalLMBatch]]: prefill = batch.past_key_values is None prefill_logprobs = batch.prefill_next_token_indices is not None @@ -579,18 +602,22 @@ class FlashCausalLM(Model): if prefill and single_request: # Ask to pre-allocate kv to its max size - # == number of tokens + max_new_tokens - pre_allocate_past_size = ( - batch.input_lengths[0] + batch.stopping_criterias[0].max_new_tokens - ) + # == Sum over batch size (number of tokens + max_new_tokens) - batch size + pre_allocate_past_size = batch.max_tokens + start_seq = batch.start_seq_prefill + end_seq = batch.end_seq_prefill else: pre_allocate_past_size = None + start_seq = batch.start_seq + end_seq = batch.end_seq out, present = self.forward( batch.input_ids, batch.position_ids, - batch.cu_seqlens, - batch.cu_seqlens_q, + start_seq, + end_seq, + batch.start_seq_q, + batch.end_seq_q, batch.max_seqlen, batch.past_key_values, pre_allocate_past_size, @@ -614,55 +641,17 @@ class FlashCausalLM(Model): # When batch == 1, we will just use the batch.input_ids values directly prefill_tokens_indices = batch.input_ids.new_zeros(len(out)) - # Create batch.cu_seqlens_q for decode - batch.cu_seqlens_q = torch.arange( - 0, len(batch) + 1, device=self.device, dtype=torch.int32 - ) + # Create batch.start_seq_q and batch.end_seq_q for decode + batch.start_seq_q = torch.arange(0, len(batch), device=self.device, dtype=torch.int32) + batch.end_seq_q = batch.start_seq_q + 1 next_position_ids = batch.position_ids.new_empty(len(batch)) + # We do not need start_seq_prefill and end_seq_prefill anymore + batch.start_seq_prefill = None + batch.end_seq_prefill = None else: prefill_logprobs = None next_position_ids = batch.position_ids - # Prepare past for next decode - if len(batch) > 1: - # Used to slice next batch past - past_indices = torch.empty( - present.shape[1], dtype=torch.int64, device=self.device - ) - batch.past_key_values = present.new_empty( - ( - present.shape[0], - present.shape[1] + len(batch.requests), - *present.shape[2:], - ) - ) - - # It is actually faster to do a whole other for loop here as the copy from present to past is fairly slow - # and will run asynchronously while we do the next for loop - cumulative_length = 0 - for i, input_length in enumerate(batch.input_lengths): - # Indexing metadata - start_index = cumulative_length - end_index = cumulative_length + input_length - - # Indices to copy present at the correct place in past_key_values - torch.arange( - start_index + i, - end_index + i, - dtype=torch.int64, - device=self.device, - out=past_indices[start_index:end_index], - ) - cumulative_length += input_length - - # Copy from present to past_key_values - batch.past_key_values[:, past_indices] = present - - # Initialize past_key_values in prefill for len(batch) == 1 - elif prefill: - # present is already pre-padded - batch.past_key_values = present - # Cumulative length cumulative_length = 0 @@ -685,6 +674,7 @@ class FlashCausalLM(Model): input_length, all_input_ids, ) in enumerate(iterator): + # Indexing metadata start_index = cumulative_length end_index = cumulative_length + input_length @@ -718,7 +708,7 @@ class FlashCausalLM(Model): # Set values in batch batch.input_ids = next_input_ids batch.position_ids = next_position_ids + 1 - batch.cu_seqlens = batch.cu_seqlens + batch.cu_seqlens_q + batch.end_seq += 1 if prefill and prefill_logprobs: # Get prefill logprobs @@ -787,7 +777,7 @@ class FlashCausalLM(Model): if stop: # Decode generated tokens output_text = self.decode( - all_input_ids[-stopping_criteria.current_tokens :] + all_input_ids[-stopping_criteria.current_tokens:] ) generated_text = GeneratedText( output_text, @@ -843,6 +833,7 @@ class FlashCausalLM(Model): batch.prefill_head_indices = None batch.prefill_next_token_indices = None batch.max_seqlen = batch.max_seqlen + 1 + batch.past_key_values = present # No need to return a batch if we know that all requests stopped return generations, batch if not stopped else None