Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Organize normal loading metadata #381

Merged
merged 2 commits into from
Jun 5, 2024
Merged

Organize normal loading metadata #381

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
741a77d
Organize normal loading metadata
EricLBuehler Jun 5, 2024
c08f7e5
Fix
EricLBuehler Jun 5, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
19 changes: 10 additions & 9 deletions mistralrs-core/src/models/gemma.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,8 +8,7 @@ use candle_nn::{linear_b as linear, Activation, RotaryEmbedding, VarBuilder};
use crate::{
device_map::DeviceMapper,
layers::{repeat_kv, CausalMasker, MatMul, QLinear, ScaledDotProductAttention},
pipeline::{extract_logits, Cache, IsqModel, NormalModel},
DeviceMapMetadata,
pipeline::{extract_logits, Cache, IsqModel, NormalLoadingMetadata, NormalModel},
};

fn default_max_position_embeddings() -> usize {
Expand Down Expand Up @@ -319,11 +318,11 @@ impl Model {
cfg: &Config,
vb: VarBuilder,
is_gptx: bool,
mapper: DeviceMapMetadata,
loading_isq: bool,
real_device: Device,
normal_loading_metadata: NormalLoadingMetadata,
) -> Result<Self> {
let mapper = mapper.into_mapper(cfg.num_hidden_layers, &real_device)?;
let mapper = normal_loading_metadata
.mapper
.into_mapper(cfg.num_hidden_layers, &normal_loading_metadata.real_device)?;
let vb_m = vb.pp("model");
let embed_tokens = candle_nn::embedding(
cfg.vocab_size,
Expand All @@ -337,7 +336,9 @@ impl Model {
cfg.rope_theta as f32,
cfg.head_dim,
cfg.max_position_embeddings,
mapper.device_for(layer_idx, false).unwrap_or(&real_device),
mapper
.device_for(layer_idx, false)
.unwrap_or(&normal_loading_metadata.real_device),
is_gptx,
vb.dtype(),
)?);
Expand All @@ -347,7 +348,7 @@ impl Model {
vb_l.pp(layer_idx),
&*mapper,
layer_idx,
loading_isq,
normal_loading_metadata.loading_isq,
)?;
layers.push(layer)
}
Expand All @@ -362,7 +363,7 @@ impl Model {
layers,
norm,
lm_head,
device: real_device,
device: normal_loading_metadata.real_device,
hidden_size: cfg.hidden_size,
cache: Cache::new(cfg.num_hidden_layers, false),
max_seq_len: default_max_position_embeddings(),
Expand Down
21 changes: 11 additions & 10 deletions mistralrs-core/src/models/llama.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,8 +10,7 @@ use std::sync::Arc;
use crate::{
device_map::DeviceMapper,
layers::{repeat_kv, CausalMasker, MatMul, RmsNorm, ScaledDotProductAttention},
pipeline::{extract_logits, IsqModel, NormalModel},
DeviceMapMetadata,
pipeline::{extract_logits, IsqModel, NormalLoadingMetadata, NormalModel},
};

#[derive(Debug, Clone, Deserialize)]
Expand Down Expand Up @@ -294,11 +293,11 @@ impl Llama {
cfg: &Config,
vb: VarBuilder,
is_gptx: bool,
mapper: DeviceMapMetadata,
loading_isq: bool,
real_device: Device,
normal_loading_metadata: NormalLoadingMetadata,
) -> Result<Self> {
let mapper = mapper.into_mapper(cfg.num_hidden_layers, &real_device)?;
let mapper = normal_loading_metadata
.mapper
.into_mapper(cfg.num_hidden_layers, &normal_loading_metadata.real_device)?;
let wte = embedding(
cfg.vocab_size,
cfg.hidden_size,
Expand All @@ -307,7 +306,7 @@ impl Llama {
let lm_head = linear(
cfg.hidden_size,
cfg.vocab_size,
mapper.set_nm_device(vb.pp("lm_head"), loading_isq),
mapper.set_nm_device(vb.pp("lm_head"), normal_loading_metadata.loading_isq),
)?;
let ln_f = RmsNorm::new(
cfg.hidden_size,
Expand All @@ -322,7 +321,9 @@ impl Llama {
cfg.rope_theta,
head_dim,
cfg.max_position_embeddings,
mapper.device_for(i, false).unwrap_or(&real_device),
mapper
.device_for(i, false)
.unwrap_or(&normal_loading_metadata.real_device),
is_gptx,
vb.dtype(),
)
Expand All @@ -333,7 +334,7 @@ impl Llama {
cfg,
&*mapper,
i,
loading_isq,
normal_loading_metadata.loading_isq,
rotary_emb,
)
.expect("Failed to load block.")
Expand All @@ -346,7 +347,7 @@ impl Llama {
ln_f,
lm_head: QMatMul::Tensor(lm_head.weight().clone()),
kv_cache: crate::pipeline::Cache::new(cfg.num_hidden_layers, false),
device: real_device,
device: normal_loading_metadata.real_device,
mapper,
})
}
Expand Down
35 changes: 13 additions & 22 deletions mistralrs-core/src/models/mistral.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,8 +8,7 @@ use std::sync::Arc;
use crate::{
device_map::DeviceMapper,
layers::{repeat_kv, CausalMasker, MatMul, RmsNorm, ScaledDotProductAttention},
pipeline::{extract_logits, Cache, IsqModel, NormalModel},
DeviceMapMetadata,
pipeline::{extract_logits, Cache, IsqModel, NormalLoadingMetadata, NormalModel},
};

#[derive(Debug, Clone, PartialEq)]
Expand Down Expand Up @@ -280,33 +279,23 @@ impl Model {
cfg: &Config,
vb: VarBuilder,
is_gptx: bool,
mapper: DeviceMapMetadata,
loading_isq: bool,
real_device: Device,
normal_loading_metadata: NormalLoadingMetadata,
) -> Result<Self> {
let vb_m = vb.pp("model");
let vb_lm_head = vb.pp("lm_head");
Self::new_inner(
cfg,
vb_m,
vb_lm_head,
is_gptx,
mapper,
loading_isq,
real_device,
)
Self::new_inner(cfg, vb_m, vb_lm_head, is_gptx, normal_loading_metadata)
}

pub fn new_inner(
cfg: &Config,
vb_m: VarBuilder,
vb_lm_head: VarBuilder,
is_gptx: bool,
mapper: DeviceMapMetadata,
loading_isq: bool,
real_device: Device,
normal_loading_metadata: NormalLoadingMetadata,
) -> Result<Self> {
let mapper = mapper.into_mapper(cfg.num_hidden_layers, &real_device)?;
let mapper = normal_loading_metadata
.mapper
.into_mapper(cfg.num_hidden_layers, &normal_loading_metadata.real_device)?;
let embed_tokens = candle_nn::embedding(
cfg.vocab_size,
cfg.hidden_size,
Expand All @@ -320,7 +309,9 @@ impl Model {
cfg.rope_theta as f32,
head_dim,
cfg.max_position_embeddings,
mapper.device_for(layer_idx, false).unwrap_or(&real_device),
mapper
.device_for(layer_idx, false)
.unwrap_or(&normal_loading_metadata.real_device),
is_gptx,
vb_m.dtype(),
)?);
Expand All @@ -330,7 +321,7 @@ impl Model {
vb_l.pp(layer_idx),
&*mapper,
layer_idx,
loading_isq,
normal_loading_metadata.loading_isq,
)?;
layers.push(layer)
}
Expand All @@ -342,15 +333,15 @@ impl Model {
let lm_head = linear_no_bias(
cfg.hidden_size,
cfg.vocab_size,
mapper.set_nm_device(vb_lm_head, loading_isq),
mapper.set_nm_device(vb_lm_head, normal_loading_metadata.loading_isq),
)?;
Ok(Self {
embed_tokens,
layers,
norm,
lm_head: QMatMul::Tensor(lm_head.weight().clone()),
sliding_window: cfg.sliding_window,
device: real_device,
device: normal_loading_metadata.real_device,
cache: Cache::new(cfg.num_hidden_layers, false),
max_seq_len: cfg.max_position_embeddings,
mapper,
Expand Down
21 changes: 11 additions & 10 deletions mistralrs-core/src/models/mixtral.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,8 +11,7 @@ use std::sync::Arc;
use crate::{
device_map::DeviceMapper,
layers::{repeat_kv, CausalMasker, MatMul, RmsNorm, ScaledDotProductAttention},
pipeline::{extract_logits, Cache, IsqModel, NormalModel},
DeviceMapMetadata,
pipeline::{extract_logits, Cache, IsqModel, NormalLoadingMetadata, NormalModel},
};

/// https://github.com/huggingface/transformers/blob/1a585c1222a56bcaecc070966d558d4a9d862e83/src/transformers/models/mixtral/configuration_mixtral.py#L113
Expand Down Expand Up @@ -383,12 +382,12 @@ impl Model {
cfg: &Config,
vb: VarBuilder,
is_gptx: bool,
mapper: DeviceMapMetadata,
loading_isq: bool,
real_device: Device,
normal_loading_metadata: NormalLoadingMetadata,
) -> Result<Self> {
let vb_m = vb.pp("model");
let mapper = mapper.into_mapper(cfg.num_hidden_layers, &real_device)?;
let mapper = normal_loading_metadata
.mapper
.into_mapper(cfg.num_hidden_layers, &normal_loading_metadata.real_device)?;
let embed_tokens = candle_nn::embedding(
cfg.vocab_size,
cfg.hidden_size,
Expand All @@ -402,7 +401,9 @@ impl Model {
cfg.rope_theta as f32,
head_dim,
cfg.max_position_embeddings,
mapper.device_for(layer_idx, false).unwrap_or(&real_device),
mapper
.device_for(layer_idx, false)
.unwrap_or(&normal_loading_metadata.real_device),
is_gptx,
vb.dtype(),
)?);
Expand All @@ -412,7 +413,7 @@ impl Model {
vb_l.pp(layer_idx),
&*mapper,
layer_idx,
loading_isq,
normal_loading_metadata.loading_isq,
)?;
layers.push(layer)
}
Expand All @@ -424,15 +425,15 @@ impl Model {
let lm_head = linear_no_bias(
cfg.hidden_size,
cfg.vocab_size,
mapper.set_nm_device(vb.pp("lm_head"), loading_isq),
mapper.set_nm_device(vb.pp("lm_head"), normal_loading_metadata.loading_isq),
)?;
Ok(Self {
embed_tokens,
layers,
norm,
lm_head: QMatMul::Tensor(lm_head.weight().clone()),
sliding_window: cfg.sliding_window,
device: real_device,
device: normal_loading_metadata.real_device,
cache: Cache::new(cfg.num_hidden_layers, false),
max_seq_len: cfg.max_position_embeddings,
mapper,
Expand Down
21 changes: 11 additions & 10 deletions mistralrs-core/src/models/phi2.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,8 +14,7 @@ use serde::Deserialize;
use crate::{
device_map::DeviceMapper,
layers::{repeat_kv, CausalMasker, QLinear, ScaledDotProductAttention},
pipeline::{extract_logits, Cache, IsqModel, NormalModel},
DeviceMapMetadata,
pipeline::{extract_logits, Cache, IsqModel, NormalLoadingMetadata, NormalModel},
};

// https://huggingface.co/microsoft/phi-2/blob/main/configuration_phi.py
Expand Down Expand Up @@ -287,12 +286,12 @@ impl Model {
cfg: &Config,
vb: VarBuilder,
is_gptx: bool,
mapper: DeviceMapMetadata,
loading_isq: bool,
real_device: Device,
normal_loading_metadata: NormalLoadingMetadata,
) -> Result<Self> {
let vb_m = vb.pp("model");
let mapper = mapper.into_mapper(cfg.num_hidden_layers, &real_device)?;
let mapper = normal_loading_metadata
.mapper
.into_mapper(cfg.num_hidden_layers, &normal_loading_metadata.real_device)?;
let embed_tokens = embedding(
cfg.vocab_size,
cfg.hidden_size,
Expand All @@ -312,7 +311,9 @@ impl Model {
cfg.head_dim(),
(cfg.partial_rotary_factor * cfg.head_dim() as f64) as usize,
cfg.max_position_embeddings,
mapper.device_for(layer_idx, false).unwrap_or(&real_device),
mapper
.device_for(layer_idx, false)
.unwrap_or(&normal_loading_metadata.real_device),
is_gptx,
vb.dtype(),
)?;
Expand All @@ -321,23 +322,23 @@ impl Model {
vb_m.pp(layer_idx),
&*mapper,
layer_idx,
loading_isq,
normal_loading_metadata.loading_isq,
rotary_emb,
)?;
layers.push(layer)
}
let lm_head = linear(
cfg.hidden_size,
cfg.vocab_size,
mapper.set_nm_device(vb.pp("lm_head"), loading_isq),
mapper.set_nm_device(vb.pp("lm_head"), normal_loading_metadata.loading_isq),
)?;
Ok(Self {
embed_tokens,
layers,
final_layernorm,
lm_head: QLinear::from_linear(lm_head),
cache: Cache::new(cfg.num_hidden_layers, false),
device: real_device,
device: normal_loading_metadata.real_device,
max_seq_len: cfg.max_position_embeddings,
mapper,
})
Expand Down
Loading
Loading