1use std::{collections::HashMap, sync::Arc};
2
3use candle_core::{DType, Device, Result, Tensor};
4use candle_nn::{Embedding, Module};
5use mistralrs_quant::{
6 ColumnParallelLayer, QuantMethod, ReplicatedLayer, RowParallelLayer, ShardedVarBuilder,
7};
8
9use crate::{
10 attention::SdpaParams,
11 device_map::DeviceMapper,
12 layers::{self, Activation, F32RmsNorm, Qwen2VLRotaryEmbedding, Sdpa},
13 paged_attention::{AttentionImplementation, ModelConfigMetadata},
14 pipeline::{
15 extract_logits, text_models_inputs_processor::FlashParams, EitherCache, IsqModel, KvCache,
16 NormalCache, NormalLoadingMetadata,
17 },
18 utils::{progress::NiceProgressBar, unvarbuilder::UnVarBuilder},
19};
20
21use super::config::Config;
22
23struct Mlp {
24 gate_proj: Arc<dyn QuantMethod>,
25 up_proj: Arc<dyn QuantMethod>,
26 down_proj: Arc<dyn QuantMethod>,
27 act_fn: Activation,
28}
29
30impl Mlp {
31 fn new(cfg: &Config, vb: ShardedVarBuilder, comm: &Arc<mistralrs_quant::Comm>) -> Result<Self> {
32 let hidden_sz = cfg.hidden_size;
33 let intermediate_sz = cfg.intermediate_size;
34 let gate_proj = ColumnParallelLayer::new(
35 hidden_sz,
36 intermediate_sz,
37 &cfg.quantization_config,
38 false,
39 comm,
40 vb.pp("gate_proj"),
41 )?;
42 let up_proj = ColumnParallelLayer::new(
43 hidden_sz,
44 intermediate_sz,
45 &cfg.quantization_config,
46 false,
47 comm,
48 vb.pp("up_proj"),
49 )?;
50 let down_proj = RowParallelLayer::new(
51 intermediate_sz,
52 hidden_sz,
53 &cfg.quantization_config,
54 false,
55 comm,
56 vb.pp("down_proj"),
57 )?;
58 Ok(Self {
59 gate_proj,
60 up_proj,
61 down_proj,
62 act_fn: cfg.hidden_act,
63 })
64 }
65
66 fn forward(&self, xs: &Tensor) -> Result<Tensor> {
67 let original_dtype = xs.dtype();
68 let mut xs = xs.clone();
69 if let Some(t) = self.gate_proj.quantized_act_type() {
70 xs = xs.to_dtype(t)?;
71 }
72 let lhs = self.gate_proj.forward(&xs)?.apply(&self.act_fn)?;
73 let rhs = self.up_proj.forward(&xs)?;
74 self.down_proj
75 .forward(&(lhs * rhs)?)?
76 .to_dtype(original_dtype)
77 }
78}
79
80struct Attention {
81 q_proj: Arc<dyn QuantMethod>,
82 k_proj: Arc<dyn QuantMethod>,
83 v_proj: Arc<dyn QuantMethod>,
84 o_proj: Arc<dyn QuantMethod>,
85 num_heads: usize,
86 num_kv_heads: usize,
87 head_dim: usize,
88 rotary_emb: Arc<Qwen2VLRotaryEmbedding>,
89 sdpa_params: SdpaParams,
90}
91
92impl Attention {
93 fn new(
94 rotary_emb: Arc<Qwen2VLRotaryEmbedding>,
95 cfg: &Config,
96 vb: ShardedVarBuilder,
97 comm: &Arc<mistralrs_quant::Comm>,
98 ) -> Result<Self> {
99 let hidden_sz = cfg.hidden_size;
100 let num_heads = cfg.num_attention_heads;
101 let num_kv_heads = cfg.num_key_value_heads;
102 let head_dim = hidden_sz / num_heads;
103 let q_proj = ColumnParallelLayer::new(
104 hidden_sz,
105 num_heads * head_dim,
106 &cfg.quantization_config,
107 true,
108 comm,
109 vb.pp("q_proj"),
110 )?;
111 let kv_shard = mistralrs_quant::compute_kv_shard(
112 cfg.num_key_value_heads,
113 cfg.hidden_size / cfg.num_attention_heads,
114 comm,
115 );
116 let k_proj = ColumnParallelLayer::new_with_shard(
117 hidden_sz,
118 num_kv_heads * head_dim,
119 &cfg.quantization_config,
120 true,
121 comm,
122 kv_shard,
123 vb.pp("k_proj"),
124 )?;
125 let v_proj = ColumnParallelLayer::new_with_shard(
126 hidden_sz,
127 num_kv_heads * head_dim,
128 &cfg.quantization_config,
129 true,
130 comm,
131 kv_shard,
132 vb.pp("v_proj"),
133 )?;
134 let o_proj = RowParallelLayer::new(
135 num_heads * head_dim,
136 hidden_sz,
137 &cfg.quantization_config,
138 false,
139 comm,
140 vb.pp("o_proj"),
141 )?;
142 Ok(Self {
143 q_proj,
144 k_proj,
145 v_proj,
146 o_proj,
147 num_heads: num_heads / comm.world_size(),
148 num_kv_heads: (num_kv_heads / comm.world_size()).max(1),
149 head_dim,
150 rotary_emb,
151 sdpa_params: SdpaParams {
152 n_kv_groups: mistralrs_quant::compute_n_kv_groups(
153 cfg.num_key_value_heads,
154 cfg.num_attention_heads,
155 comm,
156 ),
157 use_flash_attn: false,
158 softcap: None,
159 softmax_scale: 1.0 / (head_dim as f32).sqrt(),
160 sliding_window: None,
161 },
162 })
163 }
164
165 #[allow(clippy::too_many_arguments)]
166 fn forward(
167 &self,
168 xs: &Tensor,
169 attention_mask: Option<&Tensor>,
170 cos_sin: &(Tensor, Tensor),
171 kv_cache: &mut KvCache,
172 flash_params: &FlashParams,
173 ) -> Result<Tensor> {
174 let (b_sz, q_len, _) = xs.dims3()?;
175
176 let original_dtype = xs.dtype();
177 let mut xs = xs.clone();
178 if let Some(t) = self.q_proj.quantized_act_type() {
179 xs = xs.to_dtype(t)?;
180 }
181 let mut q = self.q_proj.forward(&xs)?;
182 let mut k = self.k_proj.forward(&xs)?;
183 let mut v = self.v_proj.forward(&xs)?;
184 if self.q_proj.quantized_act_type().is_some() {
185 q = q.to_dtype(original_dtype)?;
186 k = k.to_dtype(original_dtype)?;
187 v = v.to_dtype(original_dtype)?;
188 }
189
190 let (mut q, mut k, v) = if q_len != 1 {
191 let q = q
192 .reshape((b_sz, q_len, self.num_heads, self.head_dim))?
193 .transpose(1, 2)?;
194 let k = k
195 .reshape((b_sz, q_len, self.num_kv_heads, self.head_dim))?
196 .transpose(1, 2)?;
197 let v = v
198 .reshape((b_sz, q_len, self.num_kv_heads, self.head_dim))?
199 .transpose(1, 2)?;
200 (q, k, v)
201 } else {
202 let q = q.reshape((b_sz, self.num_heads, q_len, self.head_dim))?;
203 let k = k.reshape((b_sz, self.num_kv_heads, q_len, self.head_dim))?;
204 let v = v.reshape((b_sz, self.num_kv_heads, q_len, self.head_dim))?;
205 (q, k, v)
206 };
207
208 self.rotary_emb.forward(cos_sin, &mut q, &mut k)?;
209
210 let mut attn_output = {
211 let (k, v) = kv_cache.append(&k, &v)?;
212
213 Sdpa.run_attention(
214 &q.contiguous()?.to_dtype(DType::F32)?,
215 &k.contiguous()?.to_dtype(DType::F32)?,
216 &v.contiguous()?.to_dtype(DType::F32)?,
217 attention_mask
218 .map(|mask| mask.to_dtype(DType::F32).unwrap())
219 .as_ref(),
220 Some(flash_params),
221 &self.sdpa_params,
222 )?
223 .to_dtype(q.dtype())?
224 };
225
226 if let Some(t) = self.q_proj.quantized_act_type() {
227 attn_output = attn_output.to_dtype(t)?;
228 }
229 attn_output = if attention_mask.is_some() {
230 attn_output.transpose(1, 2)?.reshape((b_sz, q_len, ()))?
231 } else {
232 attn_output.reshape((b_sz, q_len, ()))?
233 };
234 let mut res = self.o_proj.forward(&attn_output)?;
235 if self.q_proj.quantized_act_type().is_some() {
236 res = res.to_dtype(original_dtype)?;
237 }
238 Ok(res)
239 }
240}
241
242pub struct DecoderLayer {
243 self_attn: Attention,
244 mlp: Mlp,
245 input_layernorm: F32RmsNorm,
246 post_attention_layernorm: F32RmsNorm,
247}
248
249impl DecoderLayer {
250 fn new(
251 rotary_emb: Arc<Qwen2VLRotaryEmbedding>,
252 cfg: &Config,
253 vb: ShardedVarBuilder,
254 mapper: &dyn DeviceMapper,
255 layer_idx: usize,
256 loading_isq: bool,
257 comm: &Arc<mistralrs_quant::Comm>,
258 ) -> Result<Self> {
259 let self_attn = Attention::new(
260 rotary_emb,
261 cfg,
262 mapper.set_device(layer_idx, vb.pp("self_attn"), loading_isq),
263 comm,
264 )?;
265 let mlp = Mlp::new(
266 cfg,
267 mapper.set_device(layer_idx, vb.pp("mlp"), loading_isq),
268 comm,
269 )?;
270 let input_layernorm = F32RmsNorm::new(
271 cfg.hidden_size,
272 cfg.rms_norm_eps,
273 mapper.set_device(layer_idx, vb.pp("input_layernorm"), false),
274 )?;
275 let post_attention_layernorm = F32RmsNorm::new(
276 cfg.hidden_size,
277 cfg.rms_norm_eps,
278 mapper.set_device(layer_idx, vb.pp("post_attention_layernorm"), false),
279 )?;
280 Ok(Self {
281 self_attn,
282 mlp,
283 input_layernorm,
284 post_attention_layernorm,
285 })
286 }
287
288 #[allow(clippy::too_many_arguments)]
289 fn forward(
290 &self,
291 xs: &Tensor,
292 attention_mask: Option<&Tensor>,
293 cos_sin: &(Tensor, Tensor),
294 kv_cache: &mut KvCache,
295 flash_params: &FlashParams,
296 ) -> Result<Tensor> {
297 let residual = xs;
298 let xs = self.input_layernorm.forward(xs)?;
299 let xs = self
300 .self_attn
301 .forward(&xs, attention_mask, cos_sin, kv_cache, flash_params)?;
302 let xs = (xs + residual)?;
303 let residual = &xs;
304 let xs = self
305 .mlp
306 .forward(&xs.apply(&self.post_attention_layernorm)?)?;
307 residual + xs
308 }
309}
310
311pub struct Qwen2VLTextModel {
312 embed_tokens: Embedding,
313 pub(super) norm: F32RmsNorm,
314 layers: Vec<DecoderLayer>,
315 mapper: Box<dyn DeviceMapper + Send + Sync>,
316 lm_head: Arc<dyn QuantMethod>,
317 pub(super) cache: EitherCache,
318 pub(super) cfg: ModelConfigMetadata,
319 pub(super) device: Device,
320 pub(super) dtype: DType,
321 pub(super) max_seq_len: usize,
322}
323
324impl Qwen2VLTextModel {
325 pub fn new(
326 cfg: &Config,
327 vb: ShardedVarBuilder,
328 _is_gptx: bool,
329 normal_loading_metadata: NormalLoadingMetadata,
330 attention_mechanism: AttentionImplementation,
331 ) -> Result<Self> {
332 if !matches!(attention_mechanism, AttentionImplementation::Eager) {
333 candle_core::bail!("Expected eager attention implementation");
334 }
335 let mapper = normal_loading_metadata.mapper;
336 let vb_m = vb.pp("model");
337
338 let embed_tokens = layers::embedding(
339 cfg.vocab_size,
340 cfg.hidden_size,
341 mapper.set_nm_device(vb_m.pp("embed_tokens"), false),
342 &cfg.quantization_config,
343 )?;
344 let mut layers = Vec::with_capacity(cfg.num_hidden_layers);
345 let head_dim = cfg.hidden_size / cfg.num_attention_heads;
346
347 let mut ropes = HashMap::new();
348 for layer_idx in 0..cfg.num_hidden_layers {
349 let device = mapper
350 .device_for(layer_idx, false)
351 .unwrap_or(&normal_loading_metadata.real_device);
352 ropes.insert(
353 device.location(),
354 Arc::new(Qwen2VLRotaryEmbedding::new(
355 cfg.rope_theta as f32,
356 head_dim,
357 device,
358 cfg.rope_scaling.mrope_section.clone(),
359 )?),
360 );
361 }
362
363 let vb_l = vb_m.pp("layers");
364 for layer_idx in NiceProgressBar::<_, 'b'>(
365 0..cfg.num_hidden_layers,
366 "Loading repeating layers",
367 &normal_loading_metadata.multi_progress,
368 ) {
369 let device = mapper
370 .device_for(layer_idx, false)
371 .unwrap_or(&normal_loading_metadata.real_device);
372 let rotary_emb = ropes
373 .get(&device.location())
374 .expect("No RoPE for device location!")
375 .clone();
376 let comm = mapper.get_comm_for(layer_idx)?;
377 let layer = DecoderLayer::new(
378 rotary_emb.clone(),
379 cfg,
380 vb_l.pp(layer_idx),
381 &*mapper,
382 layer_idx,
383 normal_loading_metadata.loading_isq,
384 &comm,
385 )?;
386 layers.push(layer)
387 }
388 let norm = F32RmsNorm::new(
389 cfg.hidden_size,
390 cfg.rms_norm_eps,
391 mapper.set_nm_device(vb_m.pp("norm"), false),
392 )?;
393 let lm_head = if !cfg.tie_word_embeddings {
394 ReplicatedLayer::new(
395 cfg.hidden_size,
396 cfg.vocab_size,
397 &None,
398 false,
399 mapper.set_nm_device(vb.pp("lm_head"), normal_loading_metadata.loading_isq),
400 )?
401 } else {
402 ReplicatedLayer::from_linear(candle_nn::Linear::new(
403 mapper.cast_nm_device(
404 embed_tokens.embeddings(),
405 normal_loading_metadata.loading_isq,
406 )?,
407 None,
408 ))?
409 };
410 Ok(Self {
411 embed_tokens,
412 norm,
413 layers,
414 lm_head,
415 cache: EitherCache::Normal(NormalCache::new(
416 cfg.num_hidden_layers,
417 cfg.max_position_embeddings,
418 )),
419 max_seq_len: cfg.max_position_embeddings,
420 cfg: ModelConfigMetadata {
421 max_seq_len: cfg.max_position_embeddings,
422 num_layers: cfg.num_hidden_layers,
423 hidden_size: cfg.hidden_size,
424 num_attn_heads: cfg.num_attention_heads / mapper.get_comm_for(0)?.world_size(),
425 num_kv_heads: (cfg.num_key_value_heads / mapper.get_comm_for(0)?.world_size())
426 .max(1),
427 sliding_window: cfg.sliding_window,
428 k_head_dim: cfg.hidden_size / cfg.num_attention_heads,
429 v_head_dim: cfg.hidden_size / cfg.num_attention_heads,
430 },
431 device: normal_loading_metadata.real_device.clone(),
432 dtype: vb.dtype(),
433 mapper,
434 })
435 }
436
437 pub fn embed_tokens(&self, input_ids: &Tensor) -> Result<Tensor> {
438 self.embed_tokens.forward(input_ids)
439 }
440
441 pub fn forward_embeds(
442 &self,
443 mut xs: Tensor,
444 attention_mask: Option<&Tensor>,
445 position_ids: &Tensor,
446 context_lens: Vec<(usize, usize)>,
447 flash_params: &FlashParams,
448 ) -> Result<Tensor> {
449 let cache = &mut self.cache.normal().0;
450 let cos_sin = self.layers[0]
451 .self_attn
452 .rotary_emb
453 .compute_cos_sin(position_ids, xs.dtype())?;
454
455 for (i, layer) in self.layers.iter().enumerate() {
456 xs = self.mapper.map(xs, i)?;
457 xs = layer.forward(
458 &xs,
459 attention_mask
460 .as_ref()
461 .map(|m| m.to_device(xs.device()).unwrap())
462 .as_ref(),
463 &cos_sin,
464 &mut cache[i],
465 flash_params,
466 )?
467 }
468 let xs = xs.to_device(&self.device)?;
469 let mut xs = xs.apply(&self.norm)?;
470 if let Some(t) = self.lm_head.quantized_act_type() {
471 xs = xs.to_dtype(t)?;
472 }
473 extract_logits(&self.lm_head.forward(&xs)?, context_lens)
474 }
475}
476
477impl IsqModel for Qwen2VLTextModel {
478 fn get_layers(
479 &mut self,
480 ) -> (
481 Vec<(&mut Arc<dyn QuantMethod>, Option<usize>)>,
482 &dyn DeviceMapper,
483 ) {
484 let mut tensors = Vec::new();
485 tensors.push((&mut self.lm_head, None));
486 for (i, layer) in self.layers.iter_mut().enumerate() {
487 tensors.push((&mut layer.self_attn.q_proj, Some(i)));
488 tensors.push((&mut layer.self_attn.k_proj, Some(i)));
489 tensors.push((&mut layer.self_attn.v_proj, Some(i)));
490 tensors.push((&mut layer.self_attn.o_proj, Some(i)));
491 tensors.push((&mut layer.mlp.gate_proj, Some(i)));
492 tensors.push((&mut layer.mlp.up_proj, Some(i)));
493 tensors.push((&mut layer.mlp.down_proj, Some(i)));
494 }
495 (tensors, &*self.mapper)
496 }
497
498 fn residual_tensors(&self) -> Vec<(String, Tensor)> {
499 let uvb = UnVarBuilder::new();
500
501 let uvb_m = uvb.pp("model");
502 uvb_m.pp("embed_tokens").add(&self.embed_tokens);
503 uvb_m.pp("norm").add(&self.norm);
504
505 for (layer_idx, layer) in self.layers.iter().enumerate() {
506 let uvb_l = uvb_m.pp("layers").pp(layer_idx);
507 uvb_l.pp("input_layernorm").add(&layer.input_layernorm);
508 uvb_l
509 .pp("post_attention_layernorm")
510 .add(&layer.post_attention_layernorm);
511 }
512
513 uvb.to_safetensors()
514 }
515}