Skip to content

GitLab

Commit a3f8d5dd authored by zhuwenwen's avatar zhuwenwen
Browse files

Merge tag 'v0.13.0rc2' into v0.13.0rc2-ori

parents 8d75f22e f34eca5f
Hide whitespace changes
Inline Side-by-side
Showing with 91 additions and 147 deletions
vllm/model_executor/models/gemma.py
View file @ a3f8d5dd
......@@ -174,7 +174,6 @@ class GemmaAttention(nn.Module):
self.rotary_emb = get_rope(
self.head_dim,
rotary_dim=self.head_dim,
max_position=max_position_embeddings,
rope_parameters=rope_parameters,
is_neox_style=True,
......
vllm/model_executor/models/gemma2.py
View file @ a3f8d5dd
......@@ -152,7 +152,6 @@ class Gemma2Attention(nn.Module):
)
self.rotary_emb = get_rope(
self.head_dim,
rotary_dim=self.head_dim,
max_position=max_position_embeddings,
rope_parameters=config.rope_parameters,
is_neox_style=True,
......
vllm/model_executor/models/gemma3.py
View file @ a3f8d5dd
......@@ -176,7 +176,6 @@ class Gemma3Attention(nn.Module):
self.rotary_emb = get_rope(
self.head_dim,
rotary_dim=self.head_dim,
max_position=max_position_embeddings,
rope_parameters=rope_parameters,
is_neox_style=True,
......
vllm/model_executor/models/gemma3_mm.py
View file @ a3f8d5dd
......@@ -237,8 +237,9 @@ class Gemma3ProcessingInfo(BaseProcessingInfo):
)
max_num_crops = images_kwargs["pan_and_scan_max_num_crops"]
# Result in the max possible feature size (h:w = max_num_crops:1)
return ImageSize(height=50 * max_num_crops, width=50)
vision_config = self.get_hf_config().vision_config
native_size = vision_config.image_size
return ImageSize(height=native_size * max_num_crops, width=native_size)
class Gemma3DummyInputsBuilder(BaseDummyInputsBuilder[Gemma3ProcessingInfo]):
......
vllm/model_executor/models/gemma3n.py
View file @ a3f8d5dd
......@@ -384,7 +384,6 @@ class Gemma3nAttention(nn.Module):
self.rotary_emb = get_rope(
self.head_dim,
rotary_dim=self.head_dim,
max_position=max_position_embeddings,
rope_parameters=rope_parameters,
is_neox_style=True,
......
vllm/model_executor/models/glm4.py
View file @ a3f8d5dd
......@@ -81,7 +81,6 @@ class Glm4Attention(nn.Module):
config.rope_parameters.setdefault("partial_rotary_factor", 0.5)
self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
self.head_dim = head_dim or hidden_size // self.total_num_heads
self.rotary_dim = self.head_dim
self.q_size = self.num_heads * self.head_dim
self.kv_size = self.num_kv_heads * self.head_dim
self.scaling = self.head_dim**-0.5
......@@ -103,7 +102,6 @@ class Glm4Attention(nn.Module):
)
self.rotary_emb = get_rope(
self.head_dim,
rotary_dim=self.rotary_dim,
max_position=max_position,
rope_parameters=config.rope_parameters,
is_neox_style=False,
......
vllm/model_executor/models/glm4_1v.py
View file @ a3f8d5dd
......@@ -47,8 +47,10 @@ from transformers.models.glm4v.video_processing_glm4v import Glm4vVideoProcessor
from transformers.video_utils import VideoMetadata
from vllm.attention.backends.registry import AttentionBackendEnum
from vllm.attention.layer import maybe_get_vit_flash_attn_backend
from vllm.config import VllmConfig
from vllm.attention.layers.mm_encoder_attention import (
MMEncoderAttention,
)
from vllm.config import MultiModalConfig, VllmConfig
from vllm.config.multimodal import BaseDummyOptions, VideoDummyOptions
from vllm.distributed import get_tensor_model_parallel_world_size, parallel_state
from vllm.distributed import utils as dist_utils
......@@ -63,6 +65,9 @@ from vllm.model_executor.layers.linear import (
)
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.rotary_embedding.common import (
ApplyRotaryEmb,
)
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
from vllm.model_executor.models.module_mapping import MultiModelKeys
from vllm.multimodal import MULTIMODAL_REGISTRY
......@@ -93,7 +98,7 @@ from .interfaces import (
SupportsMultiModal,
SupportsPP,
)
from .qwen2_vl import _create_qwen2vl_field_factory, apply_rotary_pos_emb_vision
from .qwen2_vl import _create_qwen2vl_field_factory
from .utils import (
AutoWeightsLoader,
WeightsMapper,
......@@ -191,10 +196,15 @@ class Glm4vVisionMLP(nn.Module):
hidden_features: int,
bias: bool = False,
quant_config: QuantizationConfig | None = None,
multimodal_config: MultiModalConfig | None = None,
prefix: str = "",
use_data_parallel: bool = False,
):
super().__init__()
use_data_parallel = (
multimodal_config.mm_encoder_tp_mode == "data"
if multimodal_config
else False
)
self.gate_up_proj = MergedColumnParallelLinear(
input_size=in_features,
output_sizes=[hidden_features] * 2,
......@@ -248,12 +258,16 @@ class Glm4vVisionAttention(nn.Module):
num_heads: int,
projection_size: int,
quant_config: QuantizationConfig | None = None,
multimodal_config: MultiModalConfig | None = None,
prefix: str = "",
use_data_parallel: bool = False,
attn_backend_override: AttentionBackendEnum | None = None,
) -> None:
super().__init__()
# Per attention head and per partition values.
use_data_parallel = (
multimodal_config.mm_encoder_tp_mode == "data"
if multimodal_config
else False
)
self.tp_size = (
1 if use_data_parallel else get_tensor_model_parallel_world_size()
)
......@@ -287,33 +301,13 @@ class Glm4vVisionAttention(nn.Module):
disable_tp=use_data_parallel,
)
# Detect attention implementation.
self.attn_backend = get_vit_attn_backend(
self.attn = MMEncoderAttention(
num_heads=self.num_attention_heads_per_partition,
head_size=self.hidden_size_per_attention_head,
dtype=torch.get_default_dtype(),
attn_backend_override=attn_backend_override,
multimodal_config=multimodal_config,
)
self.attn_backend, self.flash_attn_varlen_func = (
maybe_get_vit_flash_attn_backend(
self.attn_backend,
attn_backend_override=attn_backend_override,
)
)
if self.attn_backend not in {
AttentionBackendEnum.FLASH_ATTN,
AttentionBackendEnum.TORCH_SDPA,
AttentionBackendEnum.ROCM_AITER_FA,
}:
raise RuntimeError(
f"GLM-4V does not support {self.attn_backend} backend now."
)
self.is_flash_attn_backend = self.attn_backend in {
AttentionBackendEnum.FLASH_ATTN,
AttentionBackendEnum.ROCM_AITER_FA,
}
self.apply_rotary_emb = ApplyRotaryEmb(enforce_enable=True)
def split_qkv(self, qkv: torch.Tensor) -> tuple[torch.Tensor, ...]:
# [s, b, 3 * head * head_dim]
......@@ -338,61 +332,33 @@ class Glm4vVisionAttention(nn.Module):
cu_seqlens: torch.Tensor,
rotary_pos_emb_cos: torch.Tensor,
rotary_pos_emb_sin: torch.Tensor,
max_seqlen: int | None = None, # Only used for Flash Attention
max_seqlen: torch.Tensor | None = None, # Only used for Flash Attention
) -> torch.Tensor:
# [s, b, c] --> [s, b, head * 3 * head_dim]
x, _ = self.qkv(x)
# [s, b, 3 * head * head_dim] -> 3 * [s, b, head, head_dim]
q, k, v = self.split_qkv(x)
batch_size = q.shape[1]
q, k, v = (rearrange(x, "s b ... -> b s ...").contiguous() for x in (q, k, v))
if rotary_pos_emb_cos is not None and rotary_pos_emb_sin is not None:
# [2 * b, s, heads, head_dim]
qk_concat = torch.cat([q, k], dim=0)
qk_rotated = apply_rotary_pos_emb_vision(
qk_concat, rotary_pos_emb_cos, rotary_pos_emb_sin
qk_rotated = self.apply_rotary_emb(
qk_concat,
rotary_pos_emb_cos,
rotary_pos_emb_sin,
)
q, k = torch.chunk(qk_rotated, 2, dim=0)
if self.is_flash_attn_backend:
q, k, v = (rearrange(x, "b s ... -> (b s) ...") for x in [q, k, v])
output = self.flash_attn_varlen_func(
q,
k,
v,
cu_seqlens_q=cu_seqlens,
cu_seqlens_k=cu_seqlens,
max_seqlen_q=max_seqlen,
max_seqlen_k=max_seqlen,
dropout_p=0.0,
causal=False,
)
context_layer = rearrange(
output, "(b s) h d -> s b (h d)", b=batch_size
).contiguous()
elif self.attn_backend == AttentionBackendEnum.TORCH_SDPA:
# Execute attention entry by entry for speed & less VRAM.
outputs = []
lens = (cu_seqlens[1:] - cu_seqlens[:-1]).tolist()
q_chunks = torch.split(q, lens, dim=1)
k_chunks = torch.split(k, lens, dim=1)
v_chunks = torch.split(v, lens, dim=1)
for q_i, k_i, v_i in zip(q_chunks, k_chunks, v_chunks):
q_i, k_i, v_i = (
rearrange(x, "b s h d -> b h s d") for x in [q_i, k_i, v_i]
)
output_i = F.scaled_dot_product_attention(q_i, k_i, v_i, dropout_p=0.0)
output_i = rearrange(output_i, "b h s d -> b s h d ")
outputs.append(output_i)
context_layer = torch.cat(outputs, dim=1)
context_layer = rearrange(
context_layer, "b s h d -> s b (h d)"
).contiguous()
context_layer = self.attn(
query=q,
key=k,
value=v,
cu_seqlens=cu_seqlens,
max_seqlen=max_seqlen,
)
context_layer = rearrange(context_layer, "b s h d -> s b (h d)").contiguous()
output, _ = self.proj(context_layer)
return output
......@@ -406,9 +372,8 @@ class Glm4vVisionBlock(nn.Module):
mlp_hidden_dim: int,
norm_layer: Callable[[int], nn.Module] | None = None,
quant_config: QuantizationConfig | None = None,
multimodal_config: MultiModalConfig | None = None,
prefix: str = "",
use_data_parallel: bool = False,
attn_backend_override: AttentionBackendEnum | None = None,
) -> None:
super().__init__()
if norm_layer is None:
......@@ -420,17 +385,16 @@ class Glm4vVisionBlock(nn.Module):
num_heads=num_heads,
projection_size=dim,
quant_config=quant_config,
multimodal_config=multimodal_config,
prefix=f"{prefix}.attn",
use_data_parallel=use_data_parallel,
attn_backend_override=attn_backend_override,
)
self.mlp = Glm4vVisionMLP(
dim,
mlp_hidden_dim,
bias=False,
quant_config=quant_config,
multimodal_config=multimodal_config,
prefix=f"{prefix}.mlp",
use_data_parallel=use_data_parallel,
)
def forward(
......@@ -489,11 +453,16 @@ class Glm4vPatchMerger(nn.Module):
d_model: int,
context_dim: int,
quant_config: QuantizationConfig | None = None,
multimodal_config: MultiModalConfig | None = None,
bias: bool = False,
prefix: str = "",
use_data_parallel: bool = False,
) -> None:
super().__init__()
use_data_parallel = (
multimodal_config.mm_encoder_tp_mode == "data"
if multimodal_config
else False
)
self.hidden_size = d_model
self.proj = ColumnParallelLinear(
self.hidden_size,
......@@ -649,19 +618,19 @@ class Glm4vVisionTransformer(nn.Module):
vision_config: Glm4vVisionConfig,
norm_eps: float = 1e-6,
quant_config: QuantizationConfig | None = None,
multimodal_config: MultiModalConfig | None = None,
prefix: str = "",
use_data_parallel: bool = False,
attn_backend_override: AttentionBackendEnum | None = None,
) -> None:
super().__init__()
assert multimodal_config is not None, "multimodal_config must be provided"
patch_size = vision_config.patch_size
temporal_patch_size = vision_config.temporal_patch_size
in_channels = vision_config.in_channels
depth = vision_config.depth
self.hidden_size = vision_config.hidden_size
self.num_heads = vision_config.num_heads
self.use_data_parallel = use_data_parallel
self.patch_size = vision_config.patch_size
self.spatial_merge_size = vision_config.spatial_merge_size
......@@ -678,9 +647,9 @@ class Glm4vVisionTransformer(nn.Module):
head_dim = self.hidden_size // self.num_heads
self.rotary_pos_emb = get_rope(
head_size=head_dim,
rotary_dim=head_dim // 2,
max_position=8192,
is_neox_style=True,
rope_parameters={"partial_rotary_factor": 0.5},
)
self.blocks = nn.ModuleList(
[
......@@ -690,9 +659,8 @@ class Glm4vVisionTransformer(nn.Module):
mlp_hidden_dim=vision_config.out_hidden_size,
norm_layer=norm_layer,
quant_config=quant_config,
multimodal_config=multimodal_config,
prefix=f"{prefix}.blocks.{layer_idx}",
use_data_parallel=self.use_data_parallel,
attn_backend_override=attn_backend_override,
)
for layer_idx in range(depth)
]
......@@ -701,9 +669,9 @@ class Glm4vVisionTransformer(nn.Module):
d_model=vision_config.out_hidden_size,
context_dim=vision_config.intermediate_size,
quant_config=quant_config,
multimodal_config=multimodal_config,
bias=False,
prefix=f"{prefix}.merger",
use_data_parallel=self.use_data_parallel,
)
self.embeddings = Glm4vVisionEmbeddings(vision_config)
......@@ -723,7 +691,7 @@ class Glm4vVisionTransformer(nn.Module):
self.attn_backend = get_vit_attn_backend(
head_size=head_dim,
dtype=torch.get_default_dtype(),
attn_backend_override=attn_backend_override,
attn_backend_override=multimodal_config.mm_encoder_attn_backend,
)
@property
......@@ -775,13 +743,13 @@ class Glm4vVisionTransformer(nn.Module):
def compute_attn_mask_seqlen(
self,
cu_seqlens: torch.Tensor,
) -> int | None:
) -> torch.Tensor | None:
max_seqlen = None
if (
self.attn_backend == AttentionBackendEnum.FLASH_ATTN
or self.attn_backend == AttentionBackendEnum.ROCM_AITER_FA
):
max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item()
max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
return max_seqlen
def forward(
......@@ -1257,6 +1225,7 @@ class Glm4vDummyInputsBuilder(BaseDummyInputsBuilder[Glm4vProcessingInfo]):
)
height = min(height, overrides.height)
num_frames = max(num_frames, 2) # GLM 4.6V requires 2 frames
video = np.full((num_frames, width, height, 3), 255, dtype=np.uint8)
video_items = []
for i in range(num_videos):
......@@ -1464,18 +1433,12 @@ class Glm4vForConditionalGeneration(
self.multimodal_config = multimodal_config
self.use_data_parallel = multimodal_config.mm_encoder_tp_mode == "data"
attn_backend_override = (
multimodal_config.mm_encoder_attn_backend
if multimodal_config is not None
else None
)
self.visual = Glm4vVisionTransformer(
config.vision_config,
norm_eps=getattr(config, "rms_norm_eps", 1e-5),
quant_config=quant_config,
multimodal_config=multimodal_config,
prefix=maybe_prefix(prefix, "visual"),
use_data_parallel=self.use_data_parallel,
attn_backend_override=attn_backend_override,
)
if config.model_type == "glm4v":
......
vllm/model_executor/models/glm4_moe.py
View file @ a3f8d5dd
......@@ -285,7 +285,6 @@ class Glm4MoeAttention(nn.Module):
config.rope_parameters.setdefault("partial_rotary_factor", 0.5)
self.rotary_emb = get_rope(
self.head_dim,
rotary_dim=self.head_dim,
max_position=max_position_embeddings,
rope_parameters=config.rope_parameters,
)
......
vllm/model_executor/models/gpt_j.py
View file @ a3f8d5dd
......@@ -95,12 +95,13 @@ class GPTJAttention(nn.Module):
scaling = self.head_size**-0.5
assert getattr(config, "rotary", True)
assert config.rotary_dim % 2 == 0
rope_parameters = getattr(config, "rope_parameters", {})
rope_parameters["partial_rotary_factor"] = config.rotary_dim / self.head_size
max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
self.rotary_emb = get_rope(
self.head_size,
rotary_dim=config.rotary_dim,
max_position=max_position_embeddings,
rope_parameters=getattr(config, "rope_parameters", None),
rope_parameters=rope_parameters,
is_neox_style=False,
)
self.attn = Attention(
......
vllm/model_executor/models/gpt_neox.py
View file @ a3f8d5dd
......@@ -92,7 +92,6 @@ class GPTNeoXAttention(nn.Module):
max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
self.rotary_emb = get_rope(
self.head_size,
rotary_dim=self.head_size,
max_position=max_position_embeddings,
rope_parameters=config.rope_parameters,
)
......
vllm/model_executor/models/gpt_oss.py
View file @ a3f8d5dd
......@@ -67,7 +67,6 @@ class OAIAttention(nn.Module):
self.rotary_emb = get_rope(
self.head_dim,
rotary_dim=self.head_dim,
max_position=config.max_position_embeddings,
dtype=torch.float32,
rope_parameters={
......
vllm/model_executor/models/granite.py
View file @ a3f8d5dd
......@@ -160,7 +160,6 @@ class GraniteAttention(nn.Module):
self.rotary_emb = get_rope(
self.head_dim,
rotary_dim=self.head_dim,
max_position=max_position_embeddings,
rope_parameters=config.rope_parameters,
)
......
vllm/model_executor/models/granitemoe.py
View file @ a3f8d5dd
......@@ -190,7 +190,6 @@ class GraniteMoeAttention(nn.Module):
)
self.rotary_emb = get_rope(
self.head_dim,
rotary_dim=self.head_dim,
max_position=max_position,
rope_parameters=rope_parameters,
is_neox_style=True,
......
vllm/model_executor/models/granitemoehybrid.py
View file @ a3f8d5dd
......@@ -271,7 +271,6 @@ class GraniteMoeHybridAttention(nn.Module):
if config.position_embedding_type == "rope":
self.rotary_emb = get_rope(
self.head_dim,
rotary_dim=self.head_dim,
max_position=config.max_position_embeddings,
rope_parameters=config.rope_parameters,
is_neox_style=True,
......
vllm/model_executor/models/grok1.py
View file @ a3f8d5dd
......@@ -181,7 +181,6 @@ class Grok1Attention(nn.Module):
)
self.rotary_emb = get_rope(
self.head_dim,
rotary_dim=self.head_dim,
max_position=max_position,
rope_parameters=rope_parameters,
is_neox_style=True,
......
vllm/model_executor/models/hunyuan_v1.py
View file @ a3f8d5dd
......@@ -199,7 +199,6 @@ class HunYuanAttention(nn.Module):
self.rotary_emb = get_rope(
self.head_dim,
rotary_dim=self.head_dim,
max_position=max_position_embeddings,
rope_parameters=config.rope_parameters,
is_neox_style=True,
......@@ -305,7 +304,6 @@ class HunYuanCrossAttention(nn.Module):
self.rotary_emb = get_rope(
self.head_dim,
rotary_dim=self.head_dim,
max_position=max_position_embeddings,
rope_parameters=config.rope_parameters,
is_neox_style=True,
......
vllm/model_executor/models/hunyuan_vision.py
View file @ a3f8d5dd
......@@ -502,6 +502,7 @@ class HunYuanVisionTransformer(nn.Module):
cu_seqlens: list = [0]
hidden_states = x.to(device=self.device, dtype=self.dtype)
# embeddings = patch_embeds + patch_pos_embed
hidden_states = self.embeddings(hidden_states, grid_thw)
for t, h, w in grid_thw:
......@@ -515,8 +516,14 @@ class HunYuanVisionTransformer(nn.Module):
hidden_states = hidden_states.reshape(seq_len, -1)
hidden_states = hidden_states.unsqueeze(0)
for layer_num, layer in enumerate(self.layers):
hidden_states = layer(hidden_states)
# build per-image lengths once
split_lengths = [int(h) * int(w) for (_, h, w) in grid_thw]
for layer in self.layers:
# hidden_states: (1, T_total, D)
parts = hidden_states.split(split_lengths, dim=1) # list of (1, L_i, D)
parts = [layer(p) for p in parts]
hidden_states = torch.cat(parts, dim=1)
# adapter
split_lengths = (cu_seqlens[1:] - cu_seqlens[:-1]).tolist()
......
vllm/model_executor/models/interfaces.py
View file @ a3f8d5dd
......@@ -53,6 +53,22 @@ The output embeddings must be one of the following formats:
"""
def _require_is_multimodal(is_multimodal: Tensor | None) -> Tensor:
"""
A helper function to be used in the context of
[vllm.model_executor.models.interfaces.SupportsMultiModal.embed_input_ids][]
to provide a better error message.
"""
if is_multimodal is None:
raise ValueError(
"`embed_input_ids` now requires `is_multimodal` arg, "
"please update your model runner according to "
"https://github.com/vllm-project/vllm/pull/16229."
)
return is_multimodal
@runtime_checkable
class SupportsMultiModal(Protocol):
"""The interface required for all multi-modal models."""
......@@ -111,13 +127,7 @@ class SupportsMultiModal(Protocol):
the appearances of their corresponding multimodal data item in the
input prompt.
"""
if hasattr(self, "get_multimodal_embeddings"):
logger.warning_once(
"`get_multimodal_embeddings` for vLLM models is deprecated and will be "
"removed in v0.13.0 or v1.0.0, whichever is earlier. Please rename "
"this method to `embed_multimodal`."
)
return self.get_multimodal_embeddings(**kwargs)
...
def get_language_model(self) -> VllmModel:
"""
......@@ -196,17 +206,10 @@ class SupportsMultiModal(Protocol):
if multimodal_embeddings is None or len(multimodal_embeddings) == 0:
return inputs_embeds
if is_multimodal is None:
raise ValueError(
"`embed_input_ids` now requires `is_multimodal` arg, "
"please update your model runner according to "
"https://github.com/vllm-project/vllm/pull/16229."
)
return _merge_multimodal_embeddings(
inputs_embeds=inputs_embeds,
multimodal_embeddings=multimodal_embeddings,
is_multimodal=is_multimodal,
is_multimodal=_require_is_multimodal(is_multimodal),
)
......
vllm/model_executor/models/interfaces_base.py
View file @ a3f8d5dd
......@@ -49,13 +49,7 @@ class VllmModel(Protocol[T_co]):
def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
"""Apply token embeddings to `input_ids`."""
if hasattr(self, "get_input_embeddings"):
logger.warning_once(
"`get_input_embeddings` for vLLM models is deprecated and will be "
"removed in v0.13.0 or v1.0.0, whichever is earlier. Please rename "
"this method to `embed_input_ids`."
)
return self.get_input_embeddings(input_ids)
...
def forward(self, input_ids: torch.Tensor, positions: torch.Tensor) -> T_co: ...
......@@ -68,15 +62,6 @@ def _check_vllm_model_init(model: type[object] | object) -> bool:
def _check_vllm_model_embed_input_ids(model: type[object] | object) -> bool:
model_embed_input_ids = getattr(model, "embed_input_ids", None)
if not callable(model_embed_input_ids):
model_get_input_embeddings = getattr(model, "get_input_embeddings", None)
if callable(model_get_input_embeddings):
logger.warning(
"`get_input_embeddings` for vLLM models is deprecated and will be "
"removed in v0.13.0 or v1.0.0, whichever is earlier. Please rename "
"this method to `embed_input_ids`."
)
model.embed_input_ids = model_get_input_embeddings
return True
logger.warning(
"The model (%s) is missing the `embed_input_ids` method.",
model,
......
vllm/model_executor/models/internlm2.py
View file @ a3f8d5dd
......@@ -140,7 +140,6 @@ class InternLM2Attention(nn.Module):
self.rotary_emb = get_rope(
self.head_dim,
rotary_dim=self.head_dim,
max_position=max_position_embeddings,
rope_parameters=rope_parameters,
)
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment