[LLM] support quantize kv cache to fp8 (#9812)

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

@@ -37,7 +37,9 @@ except ImportError:
     rearrange = None
 
 from bigdl.llm.transformers.models.utils import extend_kv_cache, init_kv_cache, append_kv_cache
-from bigdl.llm.transformers.models.utils import rotate_half
+from bigdl.llm.transformers.models.utils import init_fp8_kv_cache, extend_fp8_kv_cache, \
+    append_fp8_kv_cache, restore_fp8_kv_cache
+from bigdl.llm.transformers.models.utils import rotate_half, quantize_kv_cache
 from bigdl.llm.utils.common import invalidInputError, invalidOperationError
 from bigdl.llm.ggml.quantize import ggml_tensor_qtype
 
@@ -83,22 +85,11 @@ def qwen_attention_forward(
     query = self._split_heads(query, self.num_heads, self.head_dim)
     key = self._split_heads(key, self.num_heads, self.head_dim)
     value = self._split_heads(value, self.num_heads, self.head_dim)
-
+    # query, key, value's shape: [bs, seq_len, num_heads, head_dim]
-    kv_seq_len = hidden_states.size()[1]
 
     if rotary_pos_emb_list is not None:
         cur_len = query.shape[1]
         if len(rotary_pos_emb_list) == 1:
-            if query.device.type == 'xpu':
-                cos, sin = rotary_pos_emb_list[0]
-                cos = cos[:, -cur_len:, :, :]
-                sin = sin[:, -cur_len:, :, :]
-                rot_dim = cos.shape[-1]
-                query_cur = query[..., :rot_dim]
-                key_cur = key[..., :rot_dim]
-                torch.ops.torch_ipex.apply_rotary_embedding(query_cur, sin, cos, query_cur)
-                torch.ops.torch_ipex.apply_rotary_embedding(key_cur, sin, cos, key_cur)
-            else:
             rotary_pos_emb = rotary_pos_emb_list[0]
             rotary_pos_emb = [i[:, -cur_len:, :, :] for i in rotary_pos_emb]
             rotary_pos_emb = (rotary_pos_emb,) * 2
@@ -119,8 +110,63 @@ def qwen_attention_forward(
             query = torch.cat(query_list, dim=0)
             key = torch.cat(key_list, dim=0)
 
-    bsz, _, n_heads, head_dim = key.size()
+    query_size, key_size = query.size(1), key.size(1)
+    kv_seq_len = key_size if layer_past is None else key_size + layer_past[0].size(1)
 
+    if kv_seq_len > self.seq_length and self.use_logn_attn and not self.training:
+        seq_start = kv_seq_len - query_size
+        seq_end = kv_seq_len
+        logn_tensor = self.logn_tensor[:, seq_start:seq_end, :, :].type_as(query)
+        query = query * logn_tensor.expand_as(query)
+    if key_size == kv_seq_len:
+        causal_mask = torch.tril(
+            torch.ones((key_size, key_size), dtype=torch.bool, device=query.device)
+        ).view(1, 1, key_size, key_size)
+    else:
+        causal_mask = None
+
+    if quantize_kv_cache(self.c_attn, hidden_states):
+        query, key, value = query.transpose(1, 2), key.transpose(1, 2), value.transpose(1, 2)
+        # query, key, value's shape: [bs, num_heads, seq_len, head_dim]
+
+        if layer_past is None:
+            # For first token, use original attn
+            attn_output, attn_weight = self._attn(
+                query, key, value, causal_mask, attention_mask, head_mask
+            )
+            if use_cache:
+                max_cache_length = kv_seq_len + KV_CACHE_ALLOC_BLOCK_LENGTH
+                k_cache, v_cache = init_fp8_kv_cache(
+                    query.size(0), self.num_heads, self.head_dim,
+                    0, max_cache_length,
+                    device=query.device,
+                )
+                key, value = append_fp8_kv_cache(k_cache, v_cache, key, value)
+        else:
+            k_cache, v_cache = layer_past[0], layer_past[1]
+            k_cache = k_cache.transpose(1, 2)
+            v_cache = v_cache.transpose(1, 2)
+            # k_cache and v_cache's shape: [bs, num_heads, context_length, head_dim]
+
+            if k_cache.stride(1) < kv_seq_len * k_cache.size(3):
+                # allocate new
+                k_cache, v_cache = extend_fp8_kv_cache(
+                    k_cache, v_cache,
+                    kv_seq_len + KV_CACHE_ALLOC_BLOCK_LENGTH,
+                    device=query.device,
+                )
+                # empty cache to reduce gpu memory
+                if v_cache.device.type == 'xpu':
+                    torch.xpu.empty_cache()
+
+            key, value = append_fp8_kv_cache(k_cache, v_cache, key, value)
+
+            attn_output, attn_weight = core_attn(
+                self, query, key, value, causal_mask, attention_mask, head_mask
+            )
+
+    else:
+        bsz = key.size(0)
         if layer_past is not None:
             cache_k, cache_v = layer_past[0], layer_past[1]
             cache_k = cache_k.transpose(1, 2)
@@ -158,23 +204,12 @@ def qwen_attention_forward(
             key = new_key_states
             value = new_value_states
 
-    query_size, key_size = query.size(1), key.size(2)
-    if key_size > self.seq_length and self.use_logn_attn and not self.training:
-        seq_start = key_size - query_size
-        seq_end = key_size
-        logn_tensor = self.logn_tensor[:, seq_start:seq_end, :, :].type_as(query)
-        query = query * logn_tensor.expand_as(query)
-    if query_size == key_size:
-        causal_mask = torch.tril(
-            torch.ones((key_size, key_size), dtype=torch.bool, device=query.device)
-        ).view(1, 1, key_size, key_size)
-    else:
-        causal_mask = None
         query = query.transpose(1, 2)
 
         attn_output, attn_weight = self._attn(
             query, key, value, causal_mask, attention_mask, head_mask
         )
+
     context_layer = self._merge_heads(
         attn_output, self.num_heads, self.head_dim
     )
@@ -191,6 +226,54 @@ def qwen_attention_forward(
     return outputs
 
 
+def core_attn(self, query, key, value, causal_mask=None, attention_mask=None, head_mask=None):
+    if query.size(2) != 1 or query.device.type != 'xpu':
+        # We have no CPU fp8 matmul implementation for now, so just upscale to fp32
+        key, value = restore_fp8_kv_cache(key, value, query.dtype)
+        attn_weights = torch.matmul(query, key.transpose(-1, -2))
+    else:
+        import linear_q4_0
+        attn_weights = linear_q4_0.query_key_fp8_matmul(query, key)
+
+    if self.scale_attn_weights:
+        if self.use_cache_quantization:
+            size_temp = value[0].size(-1)
+        else:
+            size_temp = value.size(-1)
+        attn_weights = attn_weights / (size_temp ** 0.5)
+
+    mask_value = torch.finfo(attn_weights.dtype).min
+    if causal_mask is not None:
+        attn_weights = torch.where(
+            causal_mask, attn_weights.to(attn_weights.dtype), mask_value
+        )
+
+    if attention_mask is not None:
+        attn_weights = attn_weights + attention_mask
+
+    if self.softmax_in_fp32:
+        attn_weights = torch.nn.functional.softmax(attn_weights.float(), dim=-1)
+    else:
+        attn_weights = torch.nn.functional.softmax(attn_weights, dim=-1)
+
+    attn_weights = attn_weights.type(query.dtype)
+    attn_weights = self.attn_dropout(attn_weights)
+
+    if head_mask is not None:
+        attn_weights = attn_weights * head_mask
+
+    if query.size(2) != 1 or query.device.type != 'xpu':
+        # We have no CPU fp8 matmul implementation for now, so just upscale to fp32
+        attn_output = torch.matmul(attn_weights, value)
+    else:
+        import linear_q4_0
+        attn_output = linear_q4_0.attn_value_fp8_matmul(attn_weights, value.transpose(-1, -2))
+
+    attn_output = attn_output.transpose(1, 2)
+
+    return attn_output, attn_weights
+
+
 def qwen_mlp_forward(self, x: torch.Tensor) -> torch.Tensor:
     x_2d = x.view(-1, x.shape[-1])
     if x_2d.shape[0] == 1 and x.device.type == 'xpu' \

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

@@ -14,8 +14,10 @@
 # limitations under the License.
 #
 
+import os
 import torch
 from bigdl.llm.utils.common import invalidInputError
+from bigdl.llm.ggml.quantize import ggml_tensor_qtype
 from bigdl.llm.transformers.utils import get_ipex_version
 
 
@@ -57,6 +59,66 @@ def append_kv_cache(cache_k, cache_v, key_states, value_states):
     return new_cache_k, new_cache_v
 
 
+def quantize_kv_cache(linear: torch.nn.Module, x: torch.Tensor) -> bool:
+    if os.environ.get("BIGDL_QUANTIZE_KV_CACHE", None) is not None:
+        return os.environ["BIGDL_QUANTIZE_KV_CACHE"] == "1"
+    else:
+        return x.device.type == 'xpu' and hasattr(linear, "qtype") and \
+            linear.qtype != ggml_tensor_qtype["fp16"] and linear.qtype != ggml_tensor_qtype["bf16"]
+
+
+def init_fp8_kv_cache(batch_size, num_heads, head_dim, current_length, max_length, device):
+    k_cache_storage = torch.empty(batch_size, num_heads, max_length, head_dim,
+                                  dtype=torch.uint8, device=device)
+
+    v_cache_storage = torch.empty(batch_size, num_heads, head_dim, max_length,
+                                  dtype=torch.uint8, device=device)
+
+    k_cache = k_cache_storage.as_strided((batch_size, num_heads, current_length, head_dim),
+                                         k_cache_storage.stride(), storage_offset=0)
+
+    v_cache = v_cache_storage.as_strided((batch_size, num_heads, head_dim, current_length),
+                                         v_cache_storage.stride(), storage_offset=0)
+
+    return k_cache, v_cache.transpose(-1, -2)
+
+
+def extend_fp8_kv_cache(k_cache, v_cache, max_length, device):
+    batch_size, num_heads, cur_length, head_dim = k_cache.shape
+    new_k_cache, new_v_cache = init_fp8_kv_cache(batch_size, num_heads, head_dim,
+                                                 cur_length, max_length, device)
+    new_k_cache[:] = k_cache
+    new_v_cache[:] = v_cache
+    return new_k_cache, new_v_cache
+
+
+def append_fp8_kv_cache(k_cache, v_cache, key, value):
+    batch_size, num_heads, cur_length, head_dim = k_cache.shape
+    new_length = cur_length + key.size(2)
+    new_size = (batch_size, num_heads, new_length, head_dim)
+
+    new_k_cache = k_cache.as_strided(new_size, k_cache.stride(), storage_offset=0)
+    new_v_cache = v_cache.as_strided(new_size, v_cache.stride(), storage_offset=0)
+
+    fp8_key = key.half().view(torch.uint8)[:, :, :, 1::2]
+    new_k_cache[:, :, cur_length:new_length, :] = fp8_key
+    fp8_value = value.half().view(torch.uint8)[:, :, :, 1::2]
+    new_v_cache[:, :, cur_length:new_length, :] = fp8_value
+
+    return new_k_cache, new_v_cache
+
+
+def restore_fp8_kv_cache(k_cache, v_cache, dtype):
+    new_k_cache = torch.full(k_cache.shape, 128, dtype=torch.int16, device=k_cache.device)
+    new_k_cache.view(torch.uint8)[:, :, :, 1::2] = k_cache
+    new_k_cache = new_k_cache.view(torch.half)
+    new_v_cache = torch.full(v_cache.shape, 128, dtype=torch.int16, device=v_cache.device)
+    new_v_cache.view(torch.uint8)[:, :, :, 1::2] = v_cache
+    new_v_cache = new_v_cache.view(torch.half)
+
+    return new_k_cache.to(dtype=dtype), new_v_cache.to(dtype=dtype)
+
+
 def rotate_half(x):
     """Rotates half the hidden dims of the input."""
     x1 = x[..., :x.shape[-1] // 2]