LLM: add quantize kv cache for llama. (#10086)

* feat: add quantize kv cache for llama.

* fix style.

* add quantized attention forward function.

* revert style.

* fix style.

* fix style.

* update quantized kv cache and add quantize_qkv

* fix style.

* fix style.

* optimize quantize kv cache.

* fix style.

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

@@ -40,6 +40,8 @@ import math
 import os
 import torch.nn.functional as F
 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, \
+    restore_fp8_kv_cache, use_quantize_kv_cache
 from bigdl.llm.transformers.models.utils import is_enough_kv_cache_room_4_31, \
     apply_rotary_pos_emb, is_enough_kv_cache_room_4_36
 from bigdl.llm.transformers.models.utils import apply_rotary_pos_emb_no_cache_xpu
@@ -224,6 +226,226 @@ def llama_attention_forward_4_31(
     use_cache: bool = False,
     padding_mask: Optional[torch.LongTensor] = None,
     **kwargs,
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+    if use_quantize_kv_cache(self.q_proj, hidden_states):
+        forward_function = llama_attention_forward_4_31_quantized
+    else:
+        forward_function = llama_attention_forward_4_31_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,
+        padding_mask=padding_mask,
+        kwargs=kwargs
+    )
+
+
+def llama_attention_forward_4_31_quantized(
+    self,
+    hidden_states: torch.Tensor,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_value: Optional[Tuple[torch.Tensor]] = None,
+    output_attentions: bool = False,
+    use_cache: bool = False,
+    padding_mask: Optional[torch.LongTensor] = None,
+    **kwargs,
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+    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_31(past_key_value, seq_len=q_len)
+    qtype = getattr(self.q_proj, "qtype", None)
+    qtype_check = qtype in [SYM_INT4, FP8E5]
+    no_tp = not self.config.pretraining_tp > 1
+    decoding_fast_path = (no_tp and qtype_check and use_fuse_rope
+                          and enough_kv_room and bsz * q_len == 1)
+
+    # single batch decoding fast path
+    # forward_qkv takes will perform QKV projection, rotary position embedding
+    # and save the key/value states to cache, then return query states and the
+    # extended key/value cache
+    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,
+                                                                         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:
+            kv_seq_len += past_key_value[0].shape[-2]
+
+        if use_fuse_rope:
+            query_states, key_states = apply_rotary_pos_emb_no_cache_xpu(query_states,
+                                                                         key_states,
+                                                                         position_ids,
+                                                                         "llama")
+        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, "llama")
+
+    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
+
+    # otherwise, use native attention
+    kv_seq_len = key_states.shape[-2]
+    if past_key_value is None:
+        attn_weights = torch.matmul(query_states,
+                                    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:
+            k_cache, v_cache = init_fp8_kv_cache(
+                bsz, self.num_key_value_heads, kv_seq_len, self.head_dim,
+                device=query_states.device
+            )
+            key_states, value_states = append_fp8_kv_cache(k_cache, v_cache,
+                                                           key_states, value_states)
+            past_key_value = (key_states, value_states)
+    else:
+        k_cache, v_cache = past_key_value
+        key_states, value_states = append_fp8_kv_cache(k_cache, v_cache,
+                                                       key_states, value_states)
+        kv_seq_len = key_states.shape[-2]
+        past_key_value = (key_states, value_states)
+
+        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)
+            # 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)
+            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)
+
+    if self.config.pretraining_tp > 1:
+        attn_output = attn_output.split(self.hidden_size // self.config.pretraining_tp, dim=2)
+        o_proj_slices = self.o_proj.weight.split(self.hidden_size // self.config.pretraining_tp,
+                                                 dim=1)
+        attn_output = sum([F.linear(attn_output[i], o_proj_slices[i])
+                           for i in range(self.config.pretraining_tp)])
+    else:
+        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 llama_attention_forward_4_31_original(
+    self,
+    hidden_states: torch.Tensor,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_value: Optional[Tuple[torch.Tensor]] = None,
+    output_attentions: bool = False,
+    use_cache: bool = False,
+    padding_mask: Optional[torch.LongTensor] = None,
+    **kwargs,
 ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
     bsz, q_len, hidden_size = hidden_states.size()
     device = hidden_states.device
@@ -333,13 +555,15 @@ def llama_attention_forward_4_31(
             cache_v = past_key_value[1]
             if not enough_kv_room:
                 # allocate new
-                new_cache_k, new_cache_v = extend_kv_cache(bsz,
+                new_cache_k, new_cache_v = extend_kv_cache(
-                                                           self.num_key_value_heads,  # Support GQA
+                    bsz,
-                                                           self.head_dim,
+                    self.num_key_value_heads,  # Support GQA
-                                                           cache_k.size(2),
+                    self.head_dim,
-                                                           kv_seq_len + KV_CACHE_ALLOC_BLOCK_LENGTH,
+                    cache_k.size(2),
-                                                           dtype=cache_k.dtype,
+                    kv_seq_len + KV_CACHE_ALLOC_BLOCK_LENGTH,
-                                                           device=device)
+                    dtype=cache_k.dtype,
+                    device=device
+                )
                 new_cache_k[:] = cache_k
                 new_cache_v[:] = cache_v
                 cache_k = new_cache_k