enable fp4 fused mlp and qkv (#10531)

* enable fp4 fused mlp and qkv

* update qwen

* update qwen2

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

@@ -41,6 +41,7 @@ from ipex_llm.transformers.models.utils import apply_rotary_pos_emb_cache_freq_x
 from ipex_llm.transformers.models.utils import mlp_fusion_check, GELU
 from ipex_llm.transformers.models.utils import is_enough_kv_cache_room_4_36, rotate_half
 from ipex_llm.transformers.low_bit_linear import SYM_INT4, FP8E5
+from ipex_llm.transformers.models.utils import decoding_fast_path_qtype_check
 
 KV_CACHE_ALLOC_BLOCK_LENGTH = 256
 
@@ -74,8 +75,8 @@ def should_use_fuse_rope(self, hidden_states, position_ids):
     return use_fuse_rope
 
 
-def use_decoding_fast_path(q_type, use_fuse_rope, enough_kv_room, bs):
+def use_decoding_fast_path(proj, use_fuse_rope, enough_kv_room, bs):
-    return q_type in [SYM_INT4, FP8E5] and \
+    return decoding_fast_path_qtype_check(proj) and \
         use_fuse_rope and enough_kv_room and bs == 1
 
 
@@ -137,7 +138,7 @@ def gemma_attention_forward(
 
     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)
-    decoding_fast_path = use_decoding_fast_path(self.q_proj.qtype,
+    decoding_fast_path = use_decoding_fast_path(self.q_proj,
                                                 use_fuse_rope,
                                                 enough_kv_room,
                                                 bsz * q_len)

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

@@ -50,7 +50,7 @@ from ipex_llm.transformers.models.utils import use_flash_attention, use_esimd_sd
 from ipex_llm.transformers.models.utils import mlp_fusion_check, fp16_fusion_check
 from transformers.modeling_outputs import BaseModelOutputWithPast
 from transformers.models.llama.modeling_llama import LlamaModel
-from ipex_llm.transformers.low_bit_linear import SYM_INT4, FP8E5, IQ2_XXS
+from ipex_llm.transformers.low_bit_linear import SYM_INT4, FP8E5, IQ2_XXS, FP4
 from ipex_llm.ggml.quantize import ggml_tensor_qtype
 from ipex_llm.utils.common import invalidInputError
 
@@ -64,6 +64,12 @@ from transformers import logging
 logger = logging.get_logger(__name__)
 
 
+def llama_decoding_fast_path_qtype_check(proj): 
+    # IQ2_XXS only can be used in Llama-like model
+    qtype = getattr(proj, "qtype", None)
+    return qtype in [SYM_INT4, FP8E5, IQ2_XXS, FP4]
+
+
 def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
     """
     This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states
@@ -329,8 +335,7 @@ def llama_attention_forward_4_31_quantized(
 
     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 = llama_decoding_fast_path_qtype_check(self.q_proj)
-    qtype_check = qtype in [SYM_INT4, FP8E5, IQ2_XXS]
     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)
@@ -463,8 +468,7 @@ def llama_attention_forward_4_31_original(
 
     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 = llama_decoding_fast_path_qtype_check(self.q_proj)
-    qtype_check = qtype in [SYM_INT4, FP8E5, IQ2_XXS]
     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)
@@ -692,8 +696,7 @@ def llama_attention_selective_batching_forward_4_31(
     # TODO: decoding fast path
     use_fuse_rope = should_use_fuse_rope(self, hidden_states, position_ids)
     enough_kv_room = past_key_value is not None and is_enough_kv_cache_room_4_31(past_key_value[0])
-    qtype = getattr(self.q_proj, "qtype", None)
+    qtype_check = llama_decoding_fast_path_qtype_check(self.q_proj)
-    qtype_check = qtype in [SYM_INT4, FP8E5, IQ2_XXS]
     no_tp = not self.config.pretraining_tp > 1
     decoding_fast_path = (no_tp and qtype_check and use_fuse_rope and
                           bsz * q_len == 1)
@@ -911,8 +914,7 @@ def llama_attention_forward_4_36_quantized(
     device = hidden_states.device
     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)
-    qtype = getattr(self.q_proj, "qtype", None)
+    qtype_check = llama_decoding_fast_path_qtype_check(self.q_proj)
-    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)
@@ -1093,8 +1095,7 @@ def llama_attention_forward_4_36_original(
 
     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)
-    qtype = getattr(self.q_proj, "qtype", None)
+    qtype_check = llama_decoding_fast_path_qtype_check(self.q_proj)
-    qtype_check = qtype in [SYM_INT4, FP8E5, IQ2_XXS]
     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)

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

@@ -53,6 +53,7 @@ from ipex_llm.transformers.models.utils import is_enough_kv_cache_room_4_31, \
     is_enough_kv_cache_room_4_36
 from ipex_llm.transformers.low_bit_linear import SYM_INT4, FP8E5, IQ2_XXS
 from ipex_llm.transformers.models.utils import use_flash_attention, use_esimd_sdp
+from ipex_llm.transformers.models.llama import llama_decoding_fast_path_qtype_check
 try:
     from transformers.cache_utils import Cache
 except ImportError:
@@ -81,8 +82,8 @@ def should_use_fuse_rope(self, hidden_states, position_ids):
     return use_fuse_rope
 
 
-def use_decoding_fast_path(q_type, use_fuse_rope, enough_kv_room, bs):
+def use_decoding_fast_path(proj, use_fuse_rope, enough_kv_room, bs):
-    return q_type in [SYM_INT4, FP8E5, IQ2_XXS] and \
+    return llama_decoding_fast_path_qtype_check(proj) and \
         use_fuse_rope and enough_kv_room and bs == 1
 
 
@@ -200,7 +201,7 @@ def mistral_attention_forward_quantized(
 
     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)
-    decoding_fast_path = use_decoding_fast_path(self.q_proj.qtype,
+    decoding_fast_path = use_decoding_fast_path(self.q_proj,
                                                 use_fuse_rope,
                                                 enough_kv_room,
                                                 bsz * q_len)
@@ -375,7 +376,7 @@ def mistral_attention_forward_original(
 
     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)
-    decoding_fast_path = use_decoding_fast_path(self.q_proj.qtype,
+    decoding_fast_path = use_decoding_fast_path(self.q_proj,
                                                 use_fuse_rope,
                                                 enough_kv_room,
                                                 bsz * q_len)
@@ -551,7 +552,7 @@ def mistral_attention_forward_4_36_quantized(
 
     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,
+    decoding_fast_path = use_decoding_fast_path(self.q_proj,
                                                 use_fuse_rope,
                                                 enough_kv_room,
                                                 bsz * q_len)
@@ -731,7 +732,7 @@ def mistral_attention_forward_4_36_original(
 
     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)
-    decoding_fast_path = use_decoding_fast_path(self.q_proj.qtype,
+    decoding_fast_path = use_decoding_fast_path(self.q_proj,
                                                 use_fuse_rope,
                                                 enough_kv_room,
                                                 bsz * q_len)

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

@@ -155,7 +155,7 @@ def mixtral_attention_forward(
 
     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)
-    decoding_fast_path = use_decoding_fast_path(self.q_proj.qtype,
+    decoding_fast_path = use_decoding_fast_path(self.q_proj,
                                                 use_fuse_rope,
                                                 enough_kv_room,
                                                 bsz * q_len)

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

@@ -43,6 +43,7 @@ from ipex_llm.transformers.models.utils import rotate_half, SILU
 from ipex_llm.transformers.models.utils import mlp_fusion_check
 from ipex_llm.transformers.models.utils import apply_rotary_pos_emb_cache_freq_xpu
 from ipex_llm.transformers.models.utils import use_flash_attention, use_esimd_sdp
+from ipex_llm.transformers.models.utils import decoding_fast_path_qtype_check
 from ipex_llm.utils.common import invalidInputError, invalidOperationError
 from ipex_llm.ggml.quantize import ggml_tensor_qtype
 from transformers.modeling_outputs import BaseModelOutputWithPast
@@ -137,7 +138,8 @@ def qwen_attention_forward_original(
     original_dtype = hidden_states.dtype
 
     use_fuse_rope = should_use_fuse_rope(self, hidden_states)
-    decoding_fast_path = (use_fuse_rope and bsz * q_len == 1)
+    qtype_check = decoding_fast_path_qtype_check(self.q_proj)
+    decoding_fast_path = (qtype_check and use_fuse_rope and bsz * q_len == 1)
     if decoding_fast_path:
         hidden_states = hidden_states.view(1, -1)
         cache_k, cache_v = layer_past[0], layer_past[1]
@@ -332,7 +334,8 @@ def qwen_attention_forward_quantized(
     device = hidden_states.device
 
     use_fuse_rope = should_use_fuse_rope(self, hidden_states)
-    # TODO: use when decoding_fast_path = (use_fuse_rope and bsz * q_len == 1)
+    # qtype_check = decoding_fast_path_qtype_check(self.q_proj)
+    # TODO: use when decoding_fast_path = (qtype_check and use_fuse_rope and bsz * q_len == 1)
     decoding_fast_path = False
     if decoding_fast_path:
         hidden_states = hidden_states.view(1, -1)

python/llm/src/ipex_llm/transformers/models/qwen2.py

@@ -57,6 +57,7 @@ from transformers.models.qwen2.modeling_qwen2 import Qwen2Model, apply_rotary_po
 from transformers.models.qwen2.modeling_qwen2 import _prepare_4d_causal_attention_mask_for_sdpa
 from transformers.models.qwen2.modeling_qwen2 import _prepare_4d_causal_attention_mask
 from transformers.modeling_outputs import BaseModelOutputWithPast
+from ipex_llm.transformers.models.utils import decoding_fast_path_qtype_check
 
 try:
     from transformers.cache_utils import Cache, DynamicCache
@@ -431,8 +432,7 @@ def qwen2_attention_forward_origin(
     device = hidden_states.device
 
     enough_kv_room = is_enough_kv_cache_room_4_36(past_key_value, self.layer_idx)
-    qtype = getattr(self.q_proj, "qtype", None)
+    qtype_check = decoding_fast_path_qtype_check(self.q_proj)
-    qtype_check = qtype in [SYM_INT4, FP8E5]
     decoding_fast_path = (qtype_check and use_fuse_rope
                           and enough_kv_room and bsz * q_len == 1)
     if decoding_fast_path:
@@ -601,8 +601,7 @@ def qwen2_sdpa_attention_forward(
     device = hidden_states.device
 
     enough_kv_room = is_enough_kv_cache_room_4_36(past_key_value, self.layer_idx)
-    qtype = getattr(self.q_proj, "qtype", None)
+    qtype_check = decoding_fast_path_qtype_check(self.q_proj)
-    qtype_check = qtype in [SYM_INT4, FP8E5]
     decoding_fast_path = (qtype_check and use_fuse_rope
                           and enough_kv_room and bsz * q_len == 1)
     if decoding_fast_path:

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

@@ -19,18 +19,20 @@ import torch
 from ipex_llm.utils.common import invalidInputError
 from ipex_llm.ggml.quantize import ggml_tensor_qtype
 from ipex_llm.transformers.utils import get_ipex_version, get_xpu_device_type
+from ipex_llm.transformers.low_bit_linear import SYM_INT4, SYM_INT8, FP8E5, IQ2_XXS, FP4, FP8E4
 
 FP8_KV_ALLOC_LENGTH = 512
-SYM_INT4 = ggml_tensor_qtype["sym_int4"]
-SYM_INT8 = ggml_tensor_qtype["sym_int8"]
-FP8E4 = ggml_tensor_qtype["fp8_e4m3"]
-FP8E5 = ggml_tensor_qtype["fp8_e5m2"]
 
 # used in fused mlp forward
 SILU = 0
 GELU = 1
 
 
+def decoding_fast_path_qtype_check(proj): 
+    qtype = getattr(proj, "qtype", None)
+    return qtype in [SYM_INT4, FP8E5, FP4]
+
+
 def init_kv_cache(batch_size, num_heads, head_dim, current_length, max_length, dtype, device):
     key_cache_storage = torch.empty(batch_size, num_heads,
                                     max_length, head_dim,
@@ -335,8 +337,7 @@ def mlp_fusion_check(x, qtype, training):
         return False
     if x.device.type != 'xpu':
         return False
-    if qtype not in [ggml_tensor_qtype["sym_int4"], ggml_tensor_qtype["fp8_e5m2"],
+    if qtype not in [SYM_INT4, FP8E5, FP4, IQ2_XXS]:
-                     ggml_tensor_qtype["gguf_iq2_xxs"]]:
         return False
     if training or x.requires_grad:
         return False

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

@@ -36,12 +36,13 @@ from ipex_llm.transformers.models.utils import init_fp8_kv_cache, append_fp8_kv_
     restore_fp8_kv_cache, use_quantize_kv_cache
 from ipex_llm.transformers.models.utils import is_enough_kv_cache_room_4_31, SILU
 from ipex_llm.transformers.low_bit_linear import SYM_INT4, FP8E5
+from ipex_llm.transformers.models.utils import decoding_fast_path_qtype_check
 
 KV_CACHE_ALLOC_BLOCK_LENGTH = 256
 
 
-def use_decoding_fast_path(q_type, use_fuse_rope, enough_kv_room, bs):
+def use_decoding_fast_path(proj, use_fuse_rope, enough_kv_room, bs):
-    return q_type in [SYM_INT4, FP8E5] and \
+    return decoding_fast_path_qtype_check(proj) and \
         use_fuse_rope and enough_kv_room and bs == 1