[LLM] Support transformers-v4.36.0 on mistral model (#9744)

* add support transformers-v4.36.0 on mistral model

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

* make the redundant implementation as utils

* fix code style

* fix

* fix style

* update with utils enough_kv_room

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

@@ -652,19 +652,34 @@ def _optimize_post(model, lightweight_bmm=False):
                             module.MistralRMSNorm,
                             llama_rms_norm_forward)
         else:
-            modeling_module_name = model.__class__.__module__
+            if version.parse(trans_version) >= version.parse("4.36.0"):
-            module = importlib.import_module(modeling_module_name)
+                modeling_module_name = model.__class__.__module__
-            from bigdl.llm.transformers.models.mistral import mistral_attention_forward
+                module = importlib.import_module(modeling_module_name)
-            convert_forward(model,
+                from bigdl.llm.transformers.models.mistral import mistral_attention_forward_4_36
-                            module.MistralAttention,
+                convert_forward(model,
-                            mistral_attention_forward
+                                module.MistralAttention,
-                            )
+                                mistral_attention_forward_4_36
-            convert_forward(model,
+                                )
-                            module.MistralRMSNorm,
+                convert_forward(model,
-                            llama_rms_norm_forward)
+                                module.MistralRMSNorm,
-            convert_forward(model,
+                                llama_rms_norm_forward)
-                            module.MistralMLP,
+                convert_forward(model,
-                            llama_mlp_forward)
+                                module.MistralMLP,
+                                llama_mlp_forward)
+            else:
+                modeling_module_name = model.__class__.__module__
+                module = importlib.import_module(modeling_module_name)
+                from bigdl.llm.transformers.models.mistral import mistral_attention_forward
+                convert_forward(model,
+                                module.MistralAttention,
+                                mistral_attention_forward
+                                )
+                convert_forward(model,
+                                module.MistralRMSNorm,
+                                llama_rms_norm_forward)
+                convert_forward(model,
+                                module.MistralMLP,
+                                llama_mlp_forward)
     elif model.config.model_type == "Yi":
         modeling_module_name = model.__class__.__module__
         module = importlib.import_module(modeling_module_name)

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

@@ -44,7 +44,8 @@ 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 apply_rotary_pos_emb,\
     apply_rotary_pos_emb_no_cache_xpu
-from bigdl.llm.transformers.models.utils import is_enough_kv_cache_room_4_31
+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
 
 KV_CACHE_ALLOC_BLOCK_LENGTH = 256
@@ -75,6 +76,46 @@ def use_decoding_fast_path(q_type, use_fuse_rope, enough_kv_room, bs):
     return q_type == SYM_INT4 and use_fuse_rope and enough_kv_room and bs == 1
 
 
+def compute_attn_outputs_weights(query_states, key_states, value_states, bsz, q_len, kv_seq_len,
+                                 num_heads, head_dim, hidden_size, attention_mask):
+    attn_weights = torch.matmul(
+        query_states,
+        key_states.transpose(2, 3)) / math.sqrt(head_dim)
+
+    if attn_weights.size() != (bsz, num_heads, q_len, kv_seq_len):
+        invalidInputError(
+            False,
+            f"Attention weights should be of size {(bsz, num_heads, q_len, kv_seq_len)},"
+            f" but is {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 attn_output.size() != (bsz, num_heads, q_len, head_dim):
+        invalidInputError(
+            f"`attn_output` should be of size {(bsz, num_heads, q_len, head_dim)},"
+            f" but is {attn_output.size()}"
+        )
+
+    attn_output = attn_output.transpose(1, 2).contiguous()
+    attn_output = attn_output.reshape(bsz, q_len, hidden_size)
+
+    return attn_output, attn_weights
+
+
 def mistral_attention_forward(
     self,
     hidden_states: torch.Tensor,
@@ -177,40 +218,141 @@ def mistral_attention_forward(
     key_states = repeat_kv(key_states, self.num_key_value_groups)
     value_states = repeat_kv(value_states, self.num_key_value_groups)
 
-    attn_weights = torch.matmul(
+    attn_output, attn_weights = compute_attn_outputs_weights(query_states, key_states, value_states,
-        query_states,
+                                                             bsz, q_len, kv_seq_len,
-        key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+                                                             self.num_heads, self.head_dim,
-
+                                                             self.hidden_size, attention_mask)
-    if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
+
-        invalidInputError(
+    attn_output = self.o_proj(attn_output)
-            False,
+
-            f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)},"
+    if not output_attentions:
-            f" but is {attn_weights.size()}"
+        attn_weights = None
-        )
+
-
+    return attn_output, attn_weights, past_key_value
-    if attention_mask is not None:
+
-        if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
+
-            invalidInputError(
+def mistral_attention_forward_4_36(
-                False,
+    self,
-                f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)},"
+    hidden_states: torch.Tensor,
-                f" but is {attention_mask.size()}"
+    attention_mask: Optional[torch.Tensor]=None,
-            )
+    position_ids: Optional[torch.LongTensor]=None,
-
+    past_key_value: Optional[Tuple[torch.Tensor]]=None,
-        attn_weights = attn_weights + attention_mask
+    output_attentions: bool=False,
-
+    use_cache: bool=False,
-    # upcast attention to fp32
+    padding_mask: Optional[torch.Tensor]=None,
-    attn_weights = nn.functional.\
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
-        softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+    bsz, q_len, _ = hidden_states.size()
-    attn_output = torch.matmul(attn_weights, value_states)
+    device = hidden_states.device
-
+
-    if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
+    use_fuse_rope = should_use_fuse_rope(self, hidden_states, position_ids)
-        invalidInputError(
+    enough_kv_room = is_enough_kv_cache_room_4_36(past_key_value, self.layer_idx)
-            f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)},"
+    decoding_fast_path = use_decoding_fast_path(self.q_proj.qtype,
-            f" but is {attn_output.size()}"
+                                                use_fuse_rope,
-        )
+                                                enough_kv_room,
-
+                                                bsz * q_len)
-    attn_output = attn_output.transpose(1, 2).contiguous()
+
-    attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+    if decoding_fast_path:
+        hidden_states = hidden_states.view(1, -1)
+
+        cache_k = past_key_value.key_cache[self.layer_idx]
+        cache_v = past_key_value.value_cache[self.layer_idx]
+
+        kv_seq_len = cache_k.shape[-2]
+
+        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,
+                                                                         cache_k, cache_v,
+                                                                         self.q_proj.weight.qtype,
+                                                                         kv_seq_len,
+                                                                         self.head_dim)
+        kv_seq_len += 1
+
+        # update past_key_value's seem_tokens and kv caches.
+        if self.layer_idx == 0:
+            past_key_value.seen_tokens = kv_seq_len
+        past_key_value.key_cache[self.layer_idx] = key_states
+        past_key_value.value_cache[self.layer_idx] = value_states
+
+    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,
+                                  "The cache structure has changed since version v4.36. "
+                                  f"If you are using {self.__class__.__name__} for "
+                                  "auto-regressive decodingwith 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 past_key_value is not None:
+            # update the number of seen tokens
+            if self.layer_idx == 0:
+                past_key_value.seen_tokens += key_states.shape[-2]
+
+            # reuse k, v, self_attention
+            # update `past_key_value` with `key_states` and `value_states` for layer `layer_idx`
+            if len(past_key_value.key_cache) <= self.layer_idx:
+                past_key_value.key_cache.append(key_states)
+                past_key_value.value_cache.append(value_states)
+            else:
+                cache_k = past_key_value.key_cache[self.layer_idx]
+                cache_v = past_key_value.value_cache[self.layer_idx]
+
+                if not enough_kv_room:
+                    # allocate new
+                    new_c_k, new_c_v = extend_kv_cache(bsz,
+                                                       self.num_key_value_heads,  # Support GQA
+                                                       self.head_dim,
+                                                       cache_k.size(2),
+                                                       kv_seq_len + KV_CACHE_ALLOC_BLOCK_LENGTH,
+                                                       dtype=cache_k.dtype,
+                                                       device=device)
+
+                    new_c_k[:] = cache_k
+                    new_c_v[:] = cache_v
+                    cache_k = new_c_k
+                    cache_v = new_c_v
+
+                key_states, value_states = append_kv_cache(cache_k, cache_v,
+                                                           key_states, value_states)
+
+                # update past_key_value
+                past_key_value.key_cache[self.layer_idx] = key_states
+                past_key_value.value_cache[self.layer_idx] = value_states
+
+    # repeat k/v heads if n_kv_heads < n_heads
+    key_states = repeat_kv(key_states, self.num_key_value_groups)
+    value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+    attn_output, attn_weights = compute_attn_outputs_weights(query_states, key_states, value_states,
+                                                             bsz, q_len, kv_seq_len,
+                                                             self.num_heads, self.head_dim,
+                                                             self.hidden_size, attention_mask)
 
     attn_output = self.o_proj(attn_output)