LLM: support quantized kv cache for Mistral in transformers >=4.36.0 (#10326)

* support quantize kv for mistral in transformers 4.36

* update mistral support.

* fix style.

python/llm/src/bigdl/llm/transformers/convert.py

@@ -1092,10 +1092,15 @@ def _optimize_post(model, lightweight_bmm=False):
                 modeling_module_name = model.__class__.__module__
                 module = importlib.import_module(modeling_module_name)
                 from bigdl.llm.transformers.models.mistral import mistral_attention_forward_4_36
+                from bigdl.llm.transformers.models.mistral import mistral_model_forward_4_36
                 convert_forward(model,
                                 module.MistralAttention,
                                 mistral_attention_forward_4_36
                                 )
+                convert_forward(model,
+                                module.MistralModel,
+                                mistral_model_forward_4_36
+                                )
                 convert_forward(model,
                                 module.MistralRMSNorm,
                                 llama_rms_norm_forward)

python/llm/src/bigdl/llm/transformers/models/llama.py

@@ -53,6 +53,10 @@ from transformers.models.llama.modeling_llama import LlamaModel
 from bigdl.llm.transformers.low_bit_linear import SYM_INT4, FP8E5, IQ2_XXS
 from bigdl.llm.ggml.quantize import ggml_tensor_qtype
 from bigdl.llm.utils.common import invalidInputError
+try:
+    from transformers.cache_utils import Cache
+except ImportError:
+    Cache = Tuple[torch.Tensor]
 
 
 def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
@@ -934,11 +938,11 @@ def llama_attention_forward_4_36(
     hidden_states: torch.Tensor,
     attention_mask: Optional[torch.Tensor] = None,
     position_ids: Optional[torch.LongTensor] = None,
-    past_key_value: Optional[Tuple[torch.Tensor]] = None,
+    past_key_value: Optional[Cache] = None,
     output_attentions: bool = False,
     use_cache: bool = False,
     **kwargs
-) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Cache]]:
     if use_quantize_kv_cache(self.q_proj, hidden_states):
         forward_function = llama_attention_forward_4_36_quantized
     else:
@@ -960,11 +964,11 @@ def llama_attention_forward_4_36_quantized(
     hidden_states: torch.Tensor,
     attention_mask: Optional[torch.Tensor] = None,
     position_ids: Optional[torch.LongTensor] = None,
-    past_key_value: Optional[Tuple[torch.Tensor]] = None,
+    past_key_value: Optional[Cache] = None,
     output_attentions: bool = False,
     use_cache: bool = False,
     **kwargs
-) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Cache]]:
     if "padding_mask" in kwargs:
         warnings.warn(
             "Passing `padding_mask` is deprecated and will be removed in v4.37. "
@@ -999,8 +1003,10 @@ def llama_attention_forward_4_36_quantized(
                                                                          position_ids,
                                                                          tmp_cache_k, tmp_cache_v,
                                                                          self.q_proj.weight.qtype,
+                                                                         self.v_proj.weight.qtype,
                                                                          0,
-                                                                         self.head_dim)
+                                                                         self.head_dim,
+                                                                         self.rotary_emb.base,)
     else:
         query_states = self.q_proj(hidden_states)
         key_states = self.k_proj(hidden_states)
@@ -1140,11 +1146,11 @@ def llama_attention_forward_4_36_original(
     hidden_states: torch.Tensor,
     attention_mask: Optional[torch.Tensor] = None,
     position_ids: Optional[torch.LongTensor] = None,
-    past_key_value: Optional[Tuple[torch.Tensor]] = None,
+    past_key_value: Optional[Cache] = None,
     output_attentions: bool = False,
     use_cache: bool = False,
     **kwargs
-) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Cache]]:
     if "padding_mask" in kwargs:
         warnings.warn(
             "Passing `padding_mask` is deprecated and will be removed in v4.37. "

python/llm/src/bigdl/llm/transformers/models/mistral.py

@@ -36,11 +36,13 @@
 # limitations under the License.
 """ PyTorch Mistral model."""
 import math
-from typing import Optional, Tuple
+from typing import List, Optional, Tuple, Union
 
 import torch
 from torch import nn
 import torch.nn.functional as F
+from transformers.modeling_outputs import BaseModelOutputWithPast
+from transformers.models.mistral.modeling_mistral import MistralModel
 from bigdl.llm.utils.common import invalidInputError
 from bigdl.llm.transformers.models.utils import init_kv_cache, extend_kv_cache, append_kv_cache
 from bigdl.llm.transformers.models.utils import init_fp8_kv_cache, append_fp8_kv_cache, \
@@ -51,7 +53,10 @@ from bigdl.llm.transformers.models.utils import is_enough_kv_cache_room_4_31, \
     is_enough_kv_cache_room_4_36
 from bigdl.llm.transformers.low_bit_linear import SYM_INT4, FP8E5, IQ2_XXS
 from bigdl.llm.transformers.models.utils import use_flash_attention, use_esimd_sdp
-
+try:
+    from transformers.cache_utils import Cache
+except ImportError:
+    Cache = Tuple[torch.Tensor]
 KV_CACHE_ALLOC_BLOCK_LENGTH = 256
 
 
@@ -121,6 +126,37 @@ def compute_attn_outputs_weights(query_states, key_states, value_states, bsz, q_
     return attn_output, attn_weights
 
 
+def mistral_model_forward_4_36(
+    self,
+    input_ids: torch.LongTensor = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[List[torch.FloatTensor]] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+) -> Union[Tuple, BaseModelOutputWithPast]:
+    from bigdl.llm.transformers.kv import DynamicFp8Cache
+    use_cache = use_cache if use_cache is not None else self.config.use_cache
+    if use_cache and use_quantize_kv_cache(self.layers[0].mlp.up_proj, input_ids):
+        if not isinstance(past_key_values, DynamicFp8Cache):
+            past_key_values = DynamicFp8Cache.from_legacy_cache(past_key_values)
+    return MistralModel.forward(
+        self=self,
+        input_ids=input_ids,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        return_dict=return_dict,
+    )
+
+
 def mistral_attention_forward(
     self,
     hidden_states: torch.Tensor,
@@ -480,11 +516,218 @@ def mistral_attention_forward_4_36(
     hidden_states: torch.Tensor,
     attention_mask: Optional[torch.Tensor]=None,
     position_ids: Optional[torch.LongTensor]=None,
-    past_key_value: Optional[Tuple[torch.Tensor]]=None,
+    past_key_value: Optional[Cache]=None,
     output_attentions: bool=False,
     use_cache: bool=False,
-    padding_mask: Optional[torch.Tensor]=None,
+    **kwargs
-) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Cache]]:
+    if use_quantize_kv_cache(self.q_proj, hidden_states):
+        forward_function = mistral_attention_forward_4_36_quantized
+    else:
+        forward_function = mistral_attention_forward_4_36_original
+    return forward_function(
+        self=self,
+        hidden_states=hidden_states,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_value=past_key_value,
+        output_attentions=output_attentions,
+        use_cache=use_cache,
+        kwargs=kwargs
+    )
+
+
+def mistral_attention_forward_4_36_quantized(
+    self,
+    hidden_states: torch.Tensor,
+    attention_mask: Optional[torch.Tensor]=None,
+    position_ids: Optional[torch.LongTensor]=None,
+    past_key_value: Optional[Cache]=None,
+    output_attentions: bool=False,
+    use_cache: bool=False,
+    **kwargs
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Cache]]:
+    bsz, q_len, hidden_size = hidden_states.size()
+    device = hidden_states.device
+    # for flash attention
+    original_dtype = hidden_states.dtype
+
+    use_fuse_rope = should_use_fuse_rope(self, hidden_states, position_ids)
+    enough_kv_room = is_enough_kv_cache_room_4_36(past_key_value, self.layer_idx, seq_len=q_len)
+    decoding_fast_path = use_decoding_fast_path(self.q_proj.qtype,
+                                                use_fuse_rope,
+                                                enough_kv_room,
+                                                bsz * q_len)
+
+    if decoding_fast_path:
+        hidden_states = hidden_states.view(1, -1)
+        tmp_cache_k, tmp_cache_v = init_kv_cache(
+            bsz,
+            self.num_key_value_heads,
+            self.head_dim,
+            0,
+            1,
+            dtype=hidden_states.dtype,
+            device=device
+        )
+        import linear_q4_0
+        query_states, key_states, value_states = linear_q4_0.forward_qkv(hidden_states,
+                                                                         self.q_proj.weight,
+                                                                         self.k_proj.weight,
+                                                                         self.v_proj.weight,
+                                                                         position_ids,
+                                                                         tmp_cache_k, tmp_cache_v,
+                                                                         self.q_proj.weight.qtype,
+                                                                         self.v_proj.weight.qtype,
+                                                                         0,
+                                                                         self.head_dim)
+    else:
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len,
+                                     self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len,
+                                         self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            if self.layer_idx is None:
+                invalidInputError(
+                    False,
+                    f"The cache structure has changed since version v4.36. "
+                    "If you are using {self.__class__.__name__} "
+                    "for auto-regressive decoding with k/v caching, "
+                    "please make sure to initialize the attention class "
+                    "with a layer index."
+                )
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+
+        if use_fuse_rope:
+            query_states, key_states = apply_rotary_pos_emb_no_cache_xpu(query_states,
+                                                                         key_states,
+                                                                         position_ids,
+                                                                         "mistral")
+        else:
+            cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+            query_states, key_states = apply_rotary_pos_emb(query_states, key_states,
+                                                            cos, sin, position_ids, "mistral")
+
+    if not self.training and not hidden_states.requires_grad:
+        fsdp_flag = use_flash_attention(query_states, key_states)
+    else:
+        fsdp_flag = False
+    if fsdp_flag:
+        attention_dtype = torch.float16  # use fp16 for flash attention
+    else:
+        attention_dtype = original_dtype
+
+    # repeat k/v heads if n_kv_heads < n_heads
+    key_states = repeat_kv(key_states, self.num_key_value_groups).to(device,
+                                                                     dtype=attention_dtype)
+    value_states = repeat_kv(value_states, self.num_key_value_groups).to(device,
+                                                                         dtype=attention_dtype)
+    kv_seq_len = key_states.shape[-2]
+    if len(past_key_value.key_cache) <= self.layer_idx:
+        attn_weights = torch.matmul(query_states.to(key_states.dtype),
+                                    key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+
+        if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
+            invalidInputError(
+                False,
+                f"Attention weights should be of size "
+                f"{(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
+                f" {attn_weights.size()}"
+            )
+
+        if attention_mask is not None:
+            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
+                invalidInputError(
+                    False,
+                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)},"
+                    f" but is {attention_mask.size()}"
+                )
+            attn_weights = attn_weights + attention_mask
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1,
+                                             dtype=torch.float32).to(query_states.dtype)
+        attn_output = torch.matmul(attn_weights, value_states)
+        if use_cache:
+            cache_kwargs = None
+            key_states, value_states = past_key_value.update(key_states, value_states,
+                                                             self.layer_idx, cache_kwargs)
+    else:
+        cache_kwargs = None  # Specific to RoPE models
+        key_states, value_states = past_key_value.update(key_states, value_states,
+                                                         self.layer_idx, cache_kwargs)
+        kv_seq_len = key_states.shape[-2]
+        if query_states.size(2) != 1 or query_states.device.type != 'xpu':
+            key_states, value_states = restore_fp8_kv_cache(key_states, value_states,
+                                                            query_states.dtype)
+            attn_weights = torch.matmul(query_states, key_states.transpose(2, 3))
+        else:
+            import linear_q4_0
+            attn_weights = linear_q4_0.query_key_fp8_matmul(query_states, key_states)
+
+        attn_weights = attn_weights / math.sqrt(self.head_dim)
+
+        if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
+            invalidInputError(
+                False,
+                f"Attention weights should be of size "
+                f"{(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
+                f" {attn_weights.size()}"
+            )
+
+        if attention_mask is not None:
+            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
+                invalidInputError(
+                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)},"
+                    f" but is {attention_mask.size()}"
+                )
+            attn_weights = attn_weights + attention_mask
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1,
+                                             dtype=torch.float32).to(query_states.dtype)
+
+        if query_states.size(2) != 1 or query_states.device.type != 'xpu':
+            attn_output = torch.matmul(attn_weights, value_states)
+        else:
+            import linear_q4_0
+            attn_output = linear_q4_0.attn_value_fp8_matmul(attn_weights,
+                                                            value_states.transpose(-1, -2))
+
+    attn_output_size = (bsz, self.num_heads, q_len, self.head_dim)
+    if attn_output.size() != attn_output_size:
+        invalidInputError(False,
+                          f"`attn_output` should be of size {attn_output_size},"
+                          f" but is {attn_output.size()}")
+
+    attn_output = attn_output.transpose(1, 2).contiguous()
+    attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+
+    attn_output = self.o_proj(attn_output)
+
+    if not output_attentions:
+        attn_weights = None
+
+    return attn_output.to(original_dtype), attn_weights, past_key_value
+
+
+def mistral_attention_forward_4_36_original(
+    self,
+    hidden_states: torch.Tensor,
+    attention_mask: Optional[torch.Tensor]=None,
+    position_ids: Optional[torch.LongTensor]=None,
+    past_key_value: Optional[Cache]=None,
+    output_attentions: bool=False,
+    use_cache: bool=False,
+    **kwargs
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Cache]]:
     bsz, q_len, hidden_size = hidden_states.size()
     device = hidden_states.device
     # for flash attention