Source code for tensorrt_llm.models.redrafter.model
# SPDX-FileCopyrightText: Copyright (c) 2022-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from collections import OrderedDict
import tensorrt as trt
from tensorrt_llm._common import default_net
from tensorrt_llm.bindings import KVCacheType
from tensorrt_llm.functional import Tensor, cast, categorical_sample
from tensorrt_llm.models import LLaMAForCausalLM
from tensorrt_llm.models.generation_mixin import GenerationMixin
from ..._utils import pad_vocab_size, str_dtype_to_trt
from .drafter import Drafter
from .redrafter_helper import (_beam_search_candidates, _beams2tree,
_process_logits_and_hidden_states)
[docs]
class ReDrafterForCausalLM(LLaMAForCausalLM):
def __init__(self, config):
super().__init__(config)
self.dtype = str_dtype_to_trt(config.dtype)
self.vocab_size = config.vocab_size
vocab_size_padded = pad_vocab_size(self.vocab_size,
config.mapping.tp_size)
self.drafter = Drafter.from_config(config, vocab_size_padded)
self.num_beams = config.redrafter_num_beams
self.beam_candidate_length = config.redrafter_draft_len_per_beam
self.beam_length = self.beam_candidate_length + 1 # including true token
self.greedy_search = config.redrafter_greedy_search
self.is_rnn = config.redrafter_is_rnn
assert self.dtype == self.drafter.dtype, f"{self.dtype} != {self.drafter.dtype}"
def _fwd_helper(self, hidden_states, lm_logits, embedding, drafter,
kwargs: dict):
'''
Must enable remove_input_padding:
hidden_states [total_tokens, H]
lm_logits [total_tokens, V]
1. process_logits: context vs gen
a. Context: just return the last hidden states, and logits/probs
b. Gen:
i. verify: use lm_logits, draft_probs, draft_indices, draft_tokens
ii. select hidden state and update probs
3. Sample token based on probs
4. Generate candidates using hidden_states, sampled token
5. Using beams, generate validation buffers, mark them as output
6. Mark all the outputs
'''
num_beams = self.num_beams
beam_length = self.beam_length
# Get the inputs needed
rand_data_sample = kwargs['rand_data_sample']
position_ids_base = kwargs['position_ids_base']
# Step 1: Process logits and hidden states
# process the base model output (verify for gen-phase)
probs, draft_input, num_accepted_tokens, \
accepted_beam_index = _process_logits_and_hidden_states(
self, lm_logits, hidden_states, kwargs)
# NOTE: num_accepted_tokens doesn't include true token so add 1 here
num_accepted_tokens = num_accepted_tokens + 1
# At this point:
# probs : [bs, V]
# hidden_states : [bs, H]
# Step 2: Sample token
next_token = categorical_sample(probs, rand_data_sample)
# Step 3: beam search
new_draft_tokens, new_draft_logits = _beam_search_candidates(
draft_input, next_token, embedding, drafter, self.num_beams,
self.beam_length, self.is_rnn)
# Step 4: tree processing
active_tokens_flattened, new_draft_token_indices, new_mask, \
new_position_offsets, packed_position_ids, next_num_gen_tokens, max_gen_token, \
total_gen_token = _beams2tree(new_draft_tokens, num_beams, beam_length,
position_ids_base + num_accepted_tokens)
# Step 5: mark all the tensors we need
num_accepted_tokens.mark_output('num_accepted_tokens')
accepted_beam_index.mark_output('accepted_beam_index')
max_gen_token.mark_output('max_gen_token')
total_gen_token.mark_output('total_gen_token')
next_num_gen_tokens.mark_output('next_spec_decoding_generation_lengths')
active_tokens_flattened.mark_output('next_flat_tokens')
new_draft_tokens.mark_output('next_draft_tokens')
new_draft_logits.mark_output('next_draft_probs')
new_draft_token_indices.mark_output('next_draft_indices')
new_mask.mark_output('spec_decoding_mask')
new_position_offsets.mark_output('next_spec_decoding_position_offsets')
packed_position_ids.mark_output('packed_position_ids')
return next_token, probs, draft_input
[docs]
def forward(self, *args, **kwargs):
"""
0. run base model, get logits, hidden_states
"""
extra_args = [
'draft_tokens',
'draft_indices',
'draft_probs',
'device_request_types',
'redrafter_inverted_temperature',
'rand_data_validation',
'rand_data_sample',
'position_ids_base',
]
use_cache = True
base_kwargs = {k: v for k, v in kwargs.items() if k not in extra_args}
if use_cache and default_net().plugin_config.paged_kv_cache is False:
lm_logits, presents, hidden_states = super().forward(
*args, **base_kwargs)
else:
lm_logits, hidden_states = super().forward(*args, **base_kwargs)
# lm_logits could be in fp32
lm_logits_cast = cast(lm_logits, self.dtype) # no-op if same type
self.register_network_output("hidden_states",
hidden_states) # debugging
new_draft_tokens, new_draft_logits, probs = self._fwd_helper(
hidden_states,
lm_logits_cast,
self.transformer.vocab_embedding,
self.drafter,
kwargs=kwargs)
return new_draft_tokens, new_draft_logits, probs
[docs]
def prepare_inputs(self, *args, **kwargs):
"""
Inputs needed:
Assuming, max_gen_tokens = 1 + nb*(bl - 1), counting true token
device_request_types: [bs]
draft_tokens: [bs, nb, bl]
draft_indices: [bs, nb, bl]
draft_probs: [bs, nb, bl-1, V]
spec_decoding_generation_lengths: [bs]
spec_decoding_position_offsets: [bs, max_gen_tokens]
spec_decoding_packed_mask: [bs, max_gen_tokens, packed_length] **
redrafter_inverted_temperature: [bs]
rand_data_sample: [bs]
rand_data_validation: [bs, nb, bl-1]
** The mask is tricky since the boolean mask will need to be
packed in runtime. So, the last dim will be:
packed_length = ceil(max_gen_tokens/32)
"""
default_range = GenerationMixin.default_range
remove_input_padding = default_net().plugin_config.remove_input_padding
use_gpt_attention_plugin = default_net(
).plugin_config.gpt_attention_plugin
use_gemm_plugin = default_net().plugin_config.gemm_plugin
paged_kv_cache = default_net().plugin_config.paged_kv_cache
max_batch_size = kwargs['max_batch_size']
assert max_batch_size is not None
bb_range = default_range(max_batch_size)
bb0_range = default_range(max_batch_size, min_range=0, opt_offset=1)
num_beam_tokens = self.num_beams * self.beam_length
max_draft_tokens = num_beam_tokens - self.num_beams # ignore the true token
max_gen_token_len = 1 + max_draft_tokens # for the true token
max_gen_token_len_range = default_range(max_gen_token_len)
bb_max_gen_token_len_range = default_range(max_gen_token_len *
max_batch_size,
min_range=0)
kwargs['speculative_decoding_draft_tokens_external'] = False
kwargs['max_draft_len'] = max_draft_tokens
kwargs['spec_decoding_is_generation_length_variable'] = True
inputs = super().prepare_inputs(*args, **kwargs)
assert inputs['spec_decoding_params'] is not None
enable_two_optimization_profiles = GenerationMixin.has_ctx_gen_opt_profiles(
use_gpt_attention_plugin=use_gpt_attention_plugin,
use_gemm_plugin=use_gemm_plugin,
remove_input_padding=remove_input_padding,
kv_cache_type=KVCacheType.PAGED
if paged_kv_cache else KVCacheType.CONTINUOUS)
if enable_two_optimization_profiles:
bb_range = [bb_range, bb_range]
bb0_range = [bb0_range, bb0_range]
max_gen_token_len_range = [
max_gen_token_len_range, max_gen_token_len_range
]
bb_max_gen_token_len_range = [
bb_max_gen_token_len_range, bb_max_gen_token_len_range
]
num_beams_range = [self.num_beams, self.num_beams]
beam_length_range = [self.beam_length, self.beam_length]
candidate_length_range = [
self.beam_candidate_length, self.beam_candidate_length
]
vocab_size_range = [self.vocab_size, self.vocab_size]
else:
bb_range = [bb_range]
bb0_range = [bb0_range]
max_gen_token_len_range = [max_gen_token_len_range]
bb_max_gen_token_len_range = [bb_max_gen_token_len_range]
num_beams_range = [self.num_beams]
beam_length_range = [self.beam_length]
candidate_length_range = [self.beam_candidate_length]
vocab_size_range = [self.vocab_size]
device_request_types = Tensor(name='device_request_types',
dtype=trt.int32,
shape=[-1],
dim_range=OrderedDict([
('batch_size', bb_range),
]))
draft_tokens = Tensor(name='draft_tokens',
dtype=trt.int32,
shape=[-1, self.num_beams, self.beam_length],
dim_range=OrderedDict([
('batch_size_wt0', bb0_range),
('num_beams', num_beams_range),
('beam_length', beam_length_range),
]))
draft_indices = Tensor(name='draft_indices',
dtype=trt.int32,
shape=[-1, self.num_beams, self.beam_length],
dim_range=OrderedDict([
('batch_size_wt0', bb0_range),
('num_beams', num_beams_range),
('beam_length', beam_length_range),
]))
draft_probs = Tensor(
name='draft_probs',
dtype=self.dtype,
shape=[-1, self.num_beams, self.beam_length - 1, self.vocab_size],
dim_range=OrderedDict([
('batch_size_wt0', bb0_range),
('num_beams', num_beams_range),
('candidate_length', candidate_length_range),
('vocab_size', vocab_size_range),
]))
redrafter_inverted_temperature = Tensor(
name='redrafter_inverted_temperature',
dtype=self.dtype,
shape=[-1],
dim_range=OrderedDict([
("batch_size", bb_range),
]))
rand_data_validation = Tensor(
name='rand_data_validation',
dtype=self.dtype,
shape=[-1, self.num_beams, self.beam_length - 1],
dim_range=OrderedDict([
('batch_size_wt0', bb0_range),
('num_beams', num_beams_range),
('candidate_length', candidate_length_range),
]))
rand_data_sample = Tensor(name='rand_data_sample',
dtype=self.dtype,
shape=[-1],
dim_range=OrderedDict([
('batch_size', bb_range),
]))
position_ids_base = Tensor(
name="position_ids_base",
dtype=trt.int32,
shape=[-1],
dim_range=OrderedDict([
("batch_size", bb_range),
]),
)
inputs[
'device_request_types'] = device_request_types # needed by process_logits
inputs['draft_tokens'] = draft_tokens
inputs['draft_indices'] = draft_indices
inputs['draft_probs'] = draft_probs
inputs[
'redrafter_inverted_temperature'] = redrafter_inverted_temperature
inputs['rand_data_validation'] = rand_data_validation
inputs['rand_data_sample'] = rand_data_sample
inputs['position_ids_base'] = position_ids_base
return inputs