stablelm fp8 kv cache (#10672)

* stablelm fp8 kvcache

* update

* fix

* change to fp8 matmul

* fix style

* fix

* fix

* meet code review

* add comment

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

@@ -633,7 +633,7 @@ def _optimize_pre(model):
                 del module.c_attn
         model.apply(split_qkv_proj_func)
     if model.config.model_type == "stablelm":
-        # For stablelm-zephyr-3b
+        # For stablelm-zephyr-3b and stablelm-2-zephyr-1_6b
         from ipex_llm.transformers.models.stablelm import merge_qkv
         model.apply(merge_qkv)
 
@@ -1342,10 +1342,11 @@ def _optimize_post(model, lightweight_bmm=False):
                         module.BertEncoder,
                         encoder_forward)
     elif model.config.model_type == 'stablelm':
-        # For stablelm-zephyr-3b
+        # For stablelm-zephyr-3b and stablelm-2-zephyr-1_6b
         modeling_module_name = model.__class__.__module__
         module = importlib.import_module(modeling_module_name)
         from ipex_llm.transformers.models.stablelm import stablelm_attention_forward
+        from ipex_llm.transformers.models.stablelm import stablelm_model_forward
         convert_forward(model,
                         module.StableLmAttention,
                         stablelm_attention_forward
@@ -1353,5 +1354,8 @@ def _optimize_post(model, lightweight_bmm=False):
         convert_forward(model,
                         module.StableLmMLP,
                         llama_mlp_forward)
-
+        convert_forward(model,
+                        module.StableLmModel,
+                        stablelm_model_forward
+                        )
     return model

python/llm/src/ipex_llm/transformers/models/stablelm.py

@@ -38,17 +38,20 @@
 #
 
 import math
-from typing import Optional, Tuple
+from typing import Optional, Tuple, List, Union
 
 import torch
 from torch import nn
 import torch.nn.functional as F
-from transformers.models.stablelm.modeling_stablelm import StableLmAttention
+from transformers.models.stablelm.modeling_stablelm import StableLmAttention, StableLmModel
+from transformers.modeling_outputs import BaseModelOutputWithPast
 
 from ipex_llm.utils.common import invalidInputError
 from ipex_llm.transformers.models.utils import extend_kv_cache, append_kv_cache
 from ipex_llm.transformers.models.utils import apply_rotary_pos_emb, \
     apply_rotary_pos_emb_cache_freq_xpu
+from ipex_llm.transformers.models.utils import init_fp8_kv_cache, append_fp8_kv_cache, \
+    restore_fp8_kv_cache, use_quantize_kv_cache
 from ipex_llm.transformers.models.utils import is_enough_kv_cache_room_4_36
 from ipex_llm.transformers.models.utils import use_flash_attention, use_esimd_sdp
 from ipex_llm.transformers.models.mistral import should_use_fuse_rope, repeat_kv
@@ -87,7 +90,68 @@ def merge_qkv(module: torch.nn.Module):
         del module.q_proj, module.k_proj, module.v_proj
 
 
+def stablelm_model_forward(
+        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 ipex_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_stablelm(self.layers[0].self_attn.head_dim,
+                                                    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 StableLmModel.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 use_quantize_kv_cache_stablelm(head_dim: int, linear: torch.nn.Module, x: torch.Tensor) -> bool:
+    return (head_dim == 64 or head_dim == 128) and use_quantize_kv_cache(linear, x)
+
+
 def stablelm_attention_forward(
+    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,
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+    if use_quantize_kv_cache_stablelm(self.head_dim, self.o_proj, hidden_states):
+        forward_function = stablelm_attention_forward_quantized
+    else:
+        forward_function = stablelm_attention_forward_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,
+    )
+
+
+def stablelm_attention_forward_original(
         self,
         hidden_states: torch.Tensor,
         attention_mask: Optional[torch.Tensor]=None,
@@ -116,8 +180,7 @@ def stablelm_attention_forward(
     kv_seq_len = key_states.shape[-2]
 
     if past_key_value is not None:
-        if self.layer_idx is None:
+        invalidInputError(self.layer_idx is not 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 "
@@ -134,6 +197,7 @@ def stablelm_attention_forward(
         key_states[..., self.rotary_emb.dim:],
     )
     cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+    # [batch_size, seq_length, num_heads, head_dim // config.partial_rotary_factor]
     if use_fuse_rope:
         query_rot, key_rot = apply_rotary_pos_emb_cache_freq_xpu(query_rot,
                                                                  key_rot,
@@ -142,7 +206,6 @@ def stablelm_attention_forward(
                                                                  "stablelm",
                                                                  position_ids)
     else:
-        # [batch_size, seq_length, num_heads, head_dim // config.partial_rotary_factor]
         query_rot, key_rot = apply_rotary_pos_emb(query_rot,
                                                   key_rot,
                                                   cos,
@@ -214,20 +277,16 @@ def stablelm_attention_forward(
             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,
+            attn_weights.size() == (bsz, self.num_heads, q_len, kv_seq_len),
             f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)},"
-                f" but is {attn_weights.size()}"
+            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,
+                attention_mask.size() == (bsz, 1, q_len, kv_seq_len),
                 f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)},"
-                    f" but is {attention_mask.size()}"
+                f" but is {attention_mask.size()}")
-                )
 
             attn_weights = attn_weights + attention_mask
 
@@ -238,12 +297,10 @@ def stablelm_attention_forward(
 
         attn_output = torch.matmul(attn_weights, value_states)
 
-        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
         invalidInputError(
-                False,
+            attn_output.size() == (bsz, self.num_heads, q_len, self.head_dim),
             f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)},"
-                f" but is {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)
@@ -253,3 +310,156 @@ def stablelm_attention_forward(
         attn_weights = None
 
     return attn_output.to(original_dtype), attn_weights, past_key_value
+
+
+def stablelm_attention_forward_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)
+    qkv = self.qkv_proj(hidden_states)
+    qkv = qkv.view(bsz, q_len, self.num_heads + 2 * self.num_key_value_heads, self.head_dim)
+    qkv = qkv.transpose(1, 2)
+    query_states, key_states, value_states = qkv.split([self.num_heads,
+                                                        self.num_heads,
+                                                        self.num_heads], dim=1)
+
+    kv_seq_len = key_states.shape[-2]
+    if past_key_value is not None:
+        invalidInputError(
+            self.layer_idx is not None,
+            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)
+
+    # Partial rotary embedding
+    query_rot, query_pass = (
+        query_states[..., : self.rotary_emb.dim],
+        query_states[..., self.rotary_emb.dim:],
+    )
+    key_rot, key_pass = (
+        key_states[..., : self.rotary_emb.dim],
+        key_states[..., self.rotary_emb.dim:],
+    )
+    cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+    # [batch_size, seq_length, num_heads, head_dim // config.partial_rotary_factor]
+    if use_fuse_rope:
+        query_rot, key_rot = apply_rotary_pos_emb_cache_freq_xpu(query_rot,
+                                                                 key_rot,
+                                                                 sin,
+                                                                 cos,
+                                                                 "stablelm",
+                                                                 position_ids)
+    else:
+        query_rot, key_rot = apply_rotary_pos_emb(query_rot,
+                                                  key_rot,
+                                                  cos,
+                                                  sin,
+                                                  position_ids,
+                                                  "stablelm")
+
+    # [batch_size, seq_length, num_heads, head_dim]
+    query_states = torch.cat((query_rot, query_pass), dim=-1)
+    key_states = torch.cat((key_rot, key_pass), dim=-1)
+
+    kv_seq_len = key_states.shape[-2]
+    if len(past_key_value.key_cache) <= self.layer_idx:
+        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3))
+        attn_weights = attn_weights / math.sqrt(self.head_dim)
+
+        invalidInputError(
+            attn_weights.size() == (bsz, self.num_heads, q_len, kv_seq_len),
+            f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}"
+            f", but is {attn_weights.size()}")
+
+        if attention_mask is not None:
+            invalidInputError(
+                attention_mask.size() == (bsz, 1, q_len, kv_seq_len),
+                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
+
+        # at inference time, for memory considerations, may not need to upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1).to(query_states.dtype)
+        attn_weights = self.attention_dropout(attn_weights)
+
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        invalidInputError(
+            attn_output.size() == (bsz, self.num_heads, q_len, self.head_dim),
+            f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}"
+            f", but is {attn_output.size()}")
+        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)
+            # 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_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)
+
+        invalidInputError(
+            attn_weights.size() == (bsz, self.num_heads, q_len, kv_seq_len),
+            f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}"
+            f", but is {attn_weights.size()}")
+
+        if attention_mask is not None:
+            invalidInputError(
+                attention_mask.size() == (bsz, 1, q_len, kv_seq_len),
+                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
+
+        # at inference time, for memory considerations, may not need to upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1)
+        attn_weights = self.attention_dropout(attn_weights)
+
+        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)
+    invalidInputError(attn_output.size() == attn_output_size,
+                      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