Skip to content

Gemma3 1b support and optimized rotating cache #1206

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

Merged
merged 3 commits into from
Mar 13, 2025
Merged

Gemma3 1b support and optimized rotating cache #1206

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
af5b72c
Support text-only gemma3
EricLBuehler Mar 13, 2025
9ab2ca9
Add rotating kv cache
EricLBuehler Mar 13, 2025
3e4f0bc
Do not preallocate rotating kv cache
EricLBuehler Mar 13, 2025
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
15 changes: 11 additions & 4 deletions mistralrs-core/src/layers.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1896,10 +1896,17 @@ impl MlpLayer for Mlp {
}
let lhs = MatMul.qmethod_matmul(&xs, &*self.gate)?;
let rhs = MatMul.qmethod_matmul(&xs, &*self.up)?;
let mut res = MatMul.qmethod_matmul(
&candle_nn::ops::mul_and_act(&lhs, &rhs, self.act.try_into()?)?,
&*self.down,
)?;
let mut res = if matches!(
self.act,
Activation::Gelu | Activation::Silu | Activation::Relu
) {
MatMul.qmethod_matmul(
&candle_nn::ops::mul_and_act(&lhs, &rhs, self.act.try_into()?)?,
&*self.down,
)?
} else {
MatMul.qmethod_matmul(&(&lhs.apply(&self.act)? * &rhs)?, &*self.down)?
};
if self.gate.quantized_act_type().is_some() {
res = res.to_dtype(original_dtype)?;
}
Expand Down
17 changes: 2 additions & 15 deletions mistralrs-core/src/layers_masker.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -161,25 +161,12 @@ impl CausalMasker {
return Ok(None);
}

let mut causal_mask = {
let causal_mask = {
let mask = self.make_mask(tgt_len, past_kv_len, input_ids.device())?;
let diagonal = past_kv_len as isize - sliding_window as isize - 1;
let context_mask = apply_tril(&mask.ones_like()?, diagonal)?;

masked_fill(&mask.to_dtype(DType::F32)?, &context_mask, f32::MIN)?
.to_dtype(DType::U8)?
};

let zero = Tensor::new(0.0f32, input_ids.device())?;
causal_mask = {
let mask = causal_mask.broadcast_as((causal_mask.dims()[0], causal_mask.dims()[1]))?;
// Mask: 1 means use from x (add 0.0), 0 means mask out (add -inf)

masked_fill(
&zero.to_dtype(dtype)?.broadcast_as(mask.shape())?,
&mask,
f32::NEG_INFINITY,
)?
masked_fill(&mask.to_dtype(DType::F32)?, &context_mask, f32::MIN)?.to_dtype(dtype)?
};

Ok(Some(causal_mask))
Expand Down
17 changes: 4 additions & 13 deletions mistralrs-core/src/models/gemma2.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -72,7 +72,6 @@ struct Attention {
head_dim: usize,
rotary_emb: Arc<RotaryEmbedding>,
use_sliding_window: bool,
sliding_window: Option<usize>,
paged_attn: Option<PagedAttention>,
sdpa_params: SdpaParams,
}
Expand Down Expand Up @@ -146,7 +145,6 @@ impl Attention {
head_dim,
rotary_emb,
use_sliding_window: layer_idx % 2 == 0, // Order is SWA, global, SWA
sliding_window,
paged_attn,
sdpa_params: SdpaParams {
n_kv_groups: mistralrs_quant::compute_n_kv_groups(
Expand Down Expand Up @@ -249,17 +247,9 @@ impl Attention {
},
None => {
// self.sliding_window is None if !self.use_sliding_window
let (k, v, mask) =
kv_cache.append_sliding_window(&k, &v, mask, self.sliding_window)?;
let (k, v) = kv_cache.append(&k, &v)?;

Sdpa.run_attention(
&q,
&k,
&v,
mask.as_ref(),
Some(flash_params),
&self.sdpa_params,
)?
Sdpa.run_attention(&q, &k, &v, mask, Some(flash_params), &self.sdpa_params)?
}
};

Expand Down Expand Up @@ -487,9 +477,10 @@ impl Model {
))?),
device: normal_loading_metadata.real_device,
hidden_size: cfg.hidden_size,
cache: EitherCache::Normal(NormalCache::new(
cache: EitherCache::Normal(NormalCache::new_sliding(
cfg.num_hidden_layers,
cfg.max_position_embeddings,
Some(cfg.sliding_window),
)),
max_seq_len: cfg.max_position_embeddings,
sliding_window: cfg.sliding_window,
Expand Down
10 changes: 4 additions & 6 deletions mistralrs-core/src/models/mistral.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -60,7 +60,6 @@ struct Attention {
num_kv_heads: usize,
head_dim: usize,
rotary_emb: Arc<RotaryEmbedding>,
sliding_window: Option<usize>,
paged_attn: Option<PagedAttention>,
sdpa_params: SdpaParams,
}
Expand Down Expand Up @@ -125,7 +124,6 @@ impl Attention {
num_kv_heads: (num_kv_heads / comm.world_size()).max(1),
head_dim,
rotary_emb,
sliding_window: cfg.sliding_window,
paged_attn,
sdpa_params: SdpaParams {
n_kv_groups: mistralrs_quant::compute_n_kv_groups(
Expand Down Expand Up @@ -220,14 +218,13 @@ impl Attention {
}
},
None => {
let (k, v, attn_mask) =
kv_cache.append_sliding_window(&k, &v, attention_mask, self.sliding_window)?;
let (k, v) = kv_cache.append(&k, &v)?;

Sdpa.run_attention(
&q,
&k,
&v,
attn_mask.as_ref(),
attention_mask,
Some(flash_params),
&self.sdpa_params,
)?
Expand Down Expand Up @@ -468,9 +465,10 @@ impl Model {
lm_head,
sliding_window: cfg.sliding_window,
device: normal_loading_metadata.real_device,
cache: EitherCache::Normal(NormalCache::new(
cache: EitherCache::Normal(NormalCache::new_sliding(
cfg.num_hidden_layers,
cfg.max_position_embeddings,
cfg.sliding_window,
)),
max_seq_len: cfg.max_position_embeddings,
cfg: ModelConfigMetadata {
Expand Down
10 changes: 4 additions & 6 deletions mistralrs-core/src/models/mixtral.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -60,7 +60,6 @@ struct Attention {
num_kv_heads: usize,
head_dim: usize,
rotary_emb: Arc<RotaryEmbedding>,
sliding_window: Option<usize>,
paged_attn: Option<PagedAttention>,
sdpa_params: SdpaParams,
}
Expand Down Expand Up @@ -125,7 +124,6 @@ impl Attention {
num_kv_heads: (num_kv_heads / comm.world_size()).max(1),
head_dim,
rotary_emb,
sliding_window: cfg.sliding_window,
paged_attn,
sdpa_params: SdpaParams {
n_kv_groups: mistralrs_quant::compute_n_kv_groups(
Expand Down Expand Up @@ -220,14 +218,13 @@ impl Attention {
}
},
None => {
let (k, v, attn_mask) =
kv_cache.append_sliding_window(&k, &v, attention_mask, self.sliding_window)?;
let (k, v) = kv_cache.append(&k, &v)?;

Sdpa.run_attention(
&q,
&k,
&v,
attn_mask.as_ref(),
attention_mask,
Some(flash_params),
&self.sdpa_params,
)?
Expand Down Expand Up @@ -605,9 +602,10 @@ impl Model {
lm_head,
sliding_window: cfg.sliding_window,
device: normal_loading_metadata.real_device,
cache: EitherCache::Normal(NormalCache::new(
cache: EitherCache::Normal(NormalCache::new_sliding(
cfg.num_hidden_layers,
cfg.max_position_embeddings,
cfg.sliding_window,
)),
max_seq_len: cfg.max_position_embeddings,
cfg: ModelConfigMetadata {
Expand Down
10 changes: 4 additions & 6 deletions mistralrs-core/src/models/phi3.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -80,7 +80,6 @@ struct Attention {
num_kv_heads: usize,
head_dim: usize,
rotary_emb: Arc<PhiRotaryEmbedding>,
sliding_window: Option<usize>,
paged_attn: Option<PagedAttention>,
sdpa_params: SdpaParams,
}
Expand Down Expand Up @@ -119,7 +118,6 @@ impl Attention {
num_heads,
num_kv_heads,
head_dim,
sliding_window: cfg.sliding_window,
paged_attn,
sdpa_params: SdpaParams {
n_kv_groups: num_heads / num_kv_heads,
Expand Down Expand Up @@ -217,14 +215,13 @@ impl Attention {
}
},
_ => {
let (k, v, attn_mask) =
kv_cache.append_sliding_window(&k, &v, attention_mask, self.sliding_window)?;
let (k, v) = kv_cache.append(&k, &v)?;

Sdpa.run_attention(
&q,
&k,
&v,
attn_mask.as_ref(),
attention_mask,
Some(flash_params),
&self.sdpa_params,
)?
Expand Down Expand Up @@ -522,9 +519,10 @@ impl Model {
norm,
lm_head,
device: normal_loading_metadata.real_device,
cache: EitherCache::Normal(NormalCache::new(
cache: EitherCache::Normal(NormalCache::new_sliding(
cfg.num_hidden_layers,
cfg.max_position_embeddings,
cfg.sliding_window,
)),
max_seq_len: cfg.max_position_embeddings,
sliding_window: cfg.sliding_window,
Expand Down
10 changes: 4 additions & 6 deletions mistralrs-core/src/models/phi3_5_moe.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -85,7 +85,6 @@ struct Attention {
num_kv_heads: usize,
head_dim: usize,
rotary_emb: Arc<PhiRotaryEmbedding>,
sliding_window: Option<usize>,
paged_attn: Option<PagedAttention>,
sdpa_params: SdpaParams,
}
Expand Down Expand Up @@ -151,7 +150,6 @@ impl Attention {
num_heads: num_heads / comm.world_size(),
num_kv_heads: (num_kv_heads / comm.world_size()).max(1),
head_dim,
sliding_window: cfg.sliding_window,
paged_attn,
sdpa_params: SdpaParams {
n_kv_groups: mistralrs_quant::compute_n_kv_groups(
Expand Down Expand Up @@ -249,14 +247,13 @@ impl Attention {
}
},
_ => {
let (k, v, attn_mask) =
kv_cache.append_sliding_window(&k, &v, attention_mask, self.sliding_window)?;
let (k, v) = kv_cache.append(&k, &v)?;

Sdpa.run_attention(
&q,
&k,
&v,
attn_mask.as_ref(),
attention_mask,
Some(flash_params),
&self.sdpa_params,
)?
Expand Down Expand Up @@ -676,9 +673,10 @@ impl Model {
norm,
lm_head,
device: normal_loading_metadata.real_device,
cache: EitherCache::Normal(NormalCache::new(
cache: EitherCache::Normal(NormalCache::new_sliding(
cfg.num_hidden_layers,
cfg.max_position_embeddings,
cfg.sliding_window,
)),
max_seq_len: cfg.max_position_embeddings,
sliding_window: cfg.sliding_window,
Expand Down
13 changes: 7 additions & 6 deletions mistralrs-core/src/models/quantized_phi3.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -54,7 +54,6 @@ struct LayerWeights {
head_dim: usize,
cos: Tensor,
sin: Tensor,
sliding_window: usize,
paged_attn: Option<PagedAttention>,
sdpa_params: SdpaParams,
dtype: DType,
Expand Down Expand Up @@ -133,10 +132,9 @@ impl LayerWeights {
)?
}
None => {
let (k, v, attn_mask) =
kv_cache.append_sliding_window(&k, &v, mask, Some(self.sliding_window))?;
let (k, v) = kv_cache.append(&k, &v)?;

Sdpa.run_attention(&q, &k, &v, attn_mask.as_ref(), None, &self.sdpa_params)?
Sdpa.run_attention(&q, &k, &v, mask, None, &self.sdpa_params)?
}
};

Expand Down Expand Up @@ -334,7 +332,6 @@ impl ModelConfig::FromGGUF for ModelWeights {
head_dim,
cos: cos.to_device(device)?,
sin: sin.to_device(device)?,
sliding_window: context_window,
paged_attn,
sdpa_params: SdpaParams {
n_kv_groups: head_count / head_count_kv,
Expand All @@ -353,7 +350,11 @@ impl ModelConfig::FromGGUF for ModelWeights {
output,
mapper: Some(mapper),
device: device.clone(),
cache: EitherCache::Normal(NormalCache::new(block_count, context_window)),
cache: EitherCache::Normal(NormalCache::new_sliding(
block_count,
context_window,
Some(context_window),
)),
max_seq_len: context_window,
dtype,
})
Expand Down
10 changes: 4 additions & 6 deletions mistralrs-core/src/models/starcoder2.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -151,7 +151,6 @@ struct Attention {
num_kv_heads: usize,
head_dim: usize,
rotary_emb: Arc<RotaryEmbedding>,
sliding_window: Option<usize>,
paged_attn: Option<PagedAttention>,
sdpa_params: SdpaParams,
}
Expand Down Expand Up @@ -217,7 +216,6 @@ impl Attention {
num_kv_heads: (num_kv_heads / comm.world_size()).max(1),
head_dim,
rotary_emb,
sliding_window: cfg.sliding_window,
paged_attn,
sdpa_params: SdpaParams {
n_kv_groups: mistralrs_quant::compute_n_kv_groups(
Expand Down Expand Up @@ -312,14 +310,13 @@ impl Attention {
}
},
None => {
let (k, v, attn_mask) =
kv_cache.append_sliding_window(&k, &v, attention_mask, self.sliding_window)?;
let (k, v) = kv_cache.append(&k, &v)?;

Sdpa.run_attention(
&q,
&k,
&v,
attn_mask.as_ref(),
attention_mask,
Some(flash_params),
&self.sdpa_params,
)?
Expand Down Expand Up @@ -526,9 +523,10 @@ impl Model {
))?),
sliding_window: cfg.sliding_window,
device: normal_loading_metadata.real_device,
cache: EitherCache::Normal(NormalCache::new(
cache: EitherCache::Normal(NormalCache::new_sliding(
cfg.num_hidden_layers,
cfg.max_position_embeddings,
cfg.sliding_window,
)),
max_seq_len: cfg.max_position_embeddings,
cfg: ModelConfigMetadata {
Expand Down
Loading
Loading