KV cache implementation for using llama models for text generation. (Comfy-Org#12195)

Author: comfyanonymous

comfy/text_encoders/llama.py

@@ -1,7 +1,7 @@
 import torch
 import torch.nn as nn
 from dataclasses import dataclass
-from typing import Optional, Any
+from typing import Optional, Any, Tuple
 import math
 
 from comfy.ldm.modules.attention import optimized_attention_for_device
@@ -32,6 +32,7 @@ class Llama2Config:
     k_norm = None
     rope_scale = None
     final_norm: bool = True
+    lm_head: bool = False
 
 @dataclass
 class Mistral3Small24BConfig:
@@ -54,6 +55,7 @@ class Mistral3Small24BConfig:
     k_norm = None
     rope_scale = None
     final_norm: bool = True
+    lm_head: bool = False
 
 @dataclass
 class Qwen25_3BConfig:
@@ -76,6 +78,7 @@ class Qwen25_3BConfig:
     k_norm = None
     rope_scale = None
     final_norm: bool = True
+    lm_head: bool = False
 
 @dataclass
 class Qwen3_06BConfig:
@@ -98,6 +101,7 @@ class Qwen3_06BConfig:
     k_norm = "gemma3"
     rope_scale = None
     final_norm: bool = True
+    lm_head: bool = False
 
 @dataclass
 class Qwen3_4BConfig:
@@ -120,6 +124,7 @@ class Qwen3_4BConfig:
     k_norm = "gemma3"
     rope_scale = None
     final_norm: bool = True
+    lm_head: bool = False
 
 @dataclass
 class Qwen3_8BConfig:
@@ -142,6 +147,7 @@ class Qwen3_8BConfig:
     k_norm = "gemma3"
     rope_scale = None
     final_norm: bool = True
+    lm_head: bool = False
 
 @dataclass
 class Ovis25_2BConfig:
@@ -164,6 +170,7 @@ class Ovis25_2BConfig:
     k_norm = "gemma3"
     rope_scale = None
     final_norm: bool = True
+    lm_head: bool = False
 
 @dataclass
 class Qwen25_7BVLI_Config:
@@ -186,6 +193,7 @@ class Qwen25_7BVLI_Config:
     k_norm = None
     rope_scale = None
     final_norm: bool = True
+    lm_head: bool = False
 
 @dataclass
 class Gemma2_2B_Config:
@@ -209,6 +217,7 @@ class Gemma2_2B_Config:
     sliding_attention = None
     rope_scale = None
     final_norm: bool = True
+    lm_head: bool = False
 
 @dataclass
 class Gemma3_4B_Config:
@@ -232,6 +241,7 @@ class Gemma3_4B_Config:
     sliding_attention = [1024, 1024, 1024, 1024, 1024, False]
     rope_scale = [8.0, 1.0]
     final_norm: bool = True
+    lm_head: bool = False
 
 @dataclass
 class Gemma3_12B_Config:
@@ -255,6 +265,7 @@ class Gemma3_12B_Config:
     sliding_attention = [1024, 1024, 1024, 1024, 1024, False]
     rope_scale = [8.0, 1.0]
     final_norm: bool = True
+    lm_head: bool = False
     vision_config = {"num_channels": 3, "hidden_act": "gelu_pytorch_tanh", "hidden_size": 1152, "image_size": 896, "intermediate_size": 4304, "model_type": "siglip_vision_model", "num_attention_heads": 16, "num_hidden_layers": 27, "patch_size": 14}
     mm_tokens_per_image = 256
 
@@ -356,6 +367,7 @@ def forward(
         attention_mask: Optional[torch.Tensor] = None,
         freqs_cis: Optional[torch.Tensor] = None,
         optimized_attention=None,
+        past_key_value: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
     ):
         batch_size, seq_length, _ = hidden_states.shape
         xq = self.q_proj(hidden_states)
@@ -373,11 +385,30 @@ def forward(
 
         xq, xk = apply_rope(xq, xk, freqs_cis=freqs_cis)
 
+        present_key_value = None
+        if past_key_value is not None:
+            index = 0
+            num_tokens = xk.shape[2]
+            if len(past_key_value) > 0:
+                past_key, past_value, index = past_key_value
+                if past_key.shape[2] >= (index + num_tokens):
+                    past_key[:, :, index:index + xk.shape[2]] = xk
+                    past_value[:, :, index:index + xv.shape[2]] = xv
+                    xk = past_key[:, :, :index + xk.shape[2]]
+                    xv = past_value[:, :, :index + xv.shape[2]]
+                    present_key_value = (past_key, past_value, index + num_tokens)
+                else:
+                    xk = torch.cat((past_key[:, :, :index], xk), dim=2)
+                    xv = torch.cat((past_value[:, :, :index], xv), dim=2)
+                    present_key_value = (xk, xv, index + num_tokens)
+            else:
+                present_key_value = (xk, xv, index + num_tokens)
+
         xk = xk.repeat_interleave(self.num_heads // self.num_kv_heads, dim=1)
         xv = xv.repeat_interleave(self.num_heads // self.num_kv_heads, dim=1)
 
         output = optimized_attention(xq, xk, xv, self.num_heads, mask=attention_mask, skip_reshape=True)
-        return self.o_proj(output)
+        return self.o_proj(output), present_key_value
 
 class MLP(nn.Module):
     def __init__(self, config: Llama2Config, device=None, dtype=None, ops: Any = None):
@@ -408,15 +439,17 @@ def forward(
         attention_mask: Optional[torch.Tensor] = None,
         freqs_cis: Optional[torch.Tensor] = None,
         optimized_attention=None,
+        past_key_value: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
     ):
         # Self Attention
         residual = x
         x = self.input_layernorm(x)
-        x = self.self_attn(
+        x, present_key_value = self.self_attn(
             hidden_states=x,
             attention_mask=attention_mask,
             freqs_cis=freqs_cis,
             optimized_attention=optimized_attention,
+            past_key_value=past_key_value,
         )
         x = residual + x
 
@@ -426,7 +459,7 @@ def forward(
         x = self.mlp(x)
         x = residual + x
 
-        return x
+        return x, present_key_value
 
 class TransformerBlockGemma2(nn.Module):
     def __init__(self, config: Llama2Config, index, device=None, dtype=None, ops: Any = None):
@@ -451,6 +484,7 @@ def forward(
         attention_mask: Optional[torch.Tensor] = None,
         freqs_cis: Optional[torch.Tensor] = None,
         optimized_attention=None,
+        past_key_value: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
     ):
         if self.transformer_type == 'gemma3':
             if self.sliding_attention:
@@ -468,11 +502,12 @@ def forward(
         # Self Attention
         residual = x
         x = self.input_layernorm(x)
-        x = self.self_attn(
+        x, present_key_value = self.self_attn(
             hidden_states=x,
             attention_mask=attention_mask,
             freqs_cis=freqs_cis,
             optimized_attention=optimized_attention,
+            past_key_value=past_key_value,
         )
 
         x = self.post_attention_layernorm(x)
@@ -485,7 +520,7 @@ def forward(
         x = self.post_feedforward_layernorm(x)
         x = residual + x
 
-        return x
+        return x, present_key_value
 
 class Llama2_(nn.Module):
     def __init__(self, config, device=None, dtype=None, ops=None):
@@ -516,9 +551,10 @@ def __init__(self, config, device=None, dtype=None, ops=None):
         else:
             self.norm = None
 
-        # self.lm_head = ops.Linear(config.hidden_size, config.vocab_size, bias=False, device=device, dtype=dtype)
+        if config.lm_head:
+            self.lm_head = ops.Linear(config.hidden_size, config.vocab_size, bias=False, device=device, dtype=dtype)
 
-    def forward(self, x, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None, position_ids=None, embeds_info=[]):
+    def forward(self, x, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None, position_ids=None, embeds_info=[], past_key_values=None):
         if embeds is not None:
             x = embeds
         else:
@@ -527,8 +563,13 @@ def forward(self, x, attention_mask=None, embeds=None, num_tokens=None, intermed
         if self.normalize_in:
             x *= self.config.hidden_size ** 0.5
 
+        seq_len = x.shape[1]
+        past_len = 0
+        if past_key_values is not None and len(past_key_values) > 0:
+            past_len = past_key_values[0][2]
+
         if position_ids is None:
-            position_ids = torch.arange(0, x.shape[1], device=x.device).unsqueeze(0)
+            position_ids = torch.arange(past_len, past_len + seq_len, device=x.device).unsqueeze(0)
 
         freqs_cis = precompute_freqs_cis(self.config.head_dim,
                                          position_ids,
@@ -539,14 +580,16 @@ def forward(self, x, attention_mask=None, embeds=None, num_tokens=None, intermed
 
         mask = None
         if attention_mask is not None:
-            mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])).expand(attention_mask.shape[0], 1, attention_mask.shape[-1], attention_mask.shape[-1])
+            mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])).expand(attention_mask.shape[0], 1, seq_len, attention_mask.shape[-1])
             mask = mask.masked_fill(mask.to(torch.bool), float("-inf"))
 
-        causal_mask = torch.empty(x.shape[1], x.shape[1], dtype=x.dtype, device=x.device).fill_(float("-inf")).triu_(1)
-        if mask is not None:
-            mask += causal_mask
-        else:
-            mask = causal_mask
+        if seq_len > 1:
+            causal_mask = torch.empty(past_len + seq_len, past_len + seq_len, dtype=x.dtype, device=x.device).fill_(float("-inf")).triu_(1)
+            if mask is not None:
+                mask += causal_mask
+            else:
+                mask = causal_mask
+
         optimized_attention = optimized_attention_for_device(x.device, mask=mask is not None, small_input=True)
 
         intermediate = None
@@ -562,16 +605,27 @@ def forward(self, x, attention_mask=None, embeds=None, num_tokens=None, intermed
             elif intermediate_output < 0:
                 intermediate_output = len(self.layers) + intermediate_output
 
+        next_key_values = []
         for i, layer in enumerate(self.layers):
             if all_intermediate is not None:
                 if only_layers is None or (i in only_layers):
                     all_intermediate.append(x.unsqueeze(1).clone())
-            x = layer(
+
+            past_kv = None
+            if past_key_values is not None:
+                past_kv = past_key_values[i] if len(past_key_values) > 0 else []
+
+            x, current_kv = layer(
                 x=x,
                 attention_mask=mask,
                 freqs_cis=freqs_cis,
                 optimized_attention=optimized_attention,
+                past_key_value=past_kv,
             )
+
+            if current_kv is not None:
+                next_key_values.append(current_kv)
+
             if i == intermediate_output:
                 intermediate = x.clone()
 
@@ -588,7 +642,10 @@ def forward(self, x, attention_mask=None, embeds=None, num_tokens=None, intermed
         if intermediate is not None and final_layer_norm_intermediate and self.norm is not None:
             intermediate = self.norm(intermediate)
 
-        return x, intermediate
+        if len(next_key_values) > 0:
+            return x, intermediate, next_key_values
+        else:
+            return x, intermediate
 
 
 class Gemma3MultiModalProjector(torch.nn.Module):