[Misc] Standardize RoPE handling for Qwen2-VL (#9250)

This commit is contained in:
Cyrus Leung
2024-10-16 13:56:17 +08:00
committed by GitHub
parent ed920135c8
commit 7e7eae338d
16 changed files with 102 additions and 200 deletions
+28 -19
View File
@@ -920,13 +920,10 @@ def get_rope(
rotary_emb = RotaryEmbedding(head_size, rotary_dim, max_position, base,
is_neox_style, dtype)
else:
scaling_type = rope_scaling[
"type"] if "type" in rope_scaling else rope_scaling["rope_type"]
# The correct one should be "longrope" but keep "su" here
# for backward compatible
if scaling_type not in {"su", "longrope"}:
scaling_factor = rope_scaling.get("factor", 1.0)
scaling_type = rope_scaling["rope_type"]
if scaling_type == "llama3":
scaling_factor = rope_scaling["factor"]
low_freq_factor = rope_scaling["low_freq_factor"]
high_freq_factor = rope_scaling["high_freq_factor"]
original_max_position = rope_scaling[
@@ -937,16 +934,39 @@ def get_rope(
scaling_factor, low_freq_factor,
high_freq_factor,
original_max_position)
elif scaling_type == "default":
if "mrope_section" in rope_scaling:
rotary_emb = MRotaryEmbedding(
head_size,
rotary_dim,
max_position,
base,
is_neox_style,
dtype,
mrope_section=rope_scaling["mrope_section"],
)
else:
rotary_emb = RotaryEmbedding(
head_size,
rotary_dim,
max_position,
base,
is_neox_style,
dtype,
)
elif scaling_type == "linear":
scaling_factor = rope_scaling["factor"]
rotary_emb = LinearScalingRotaryEmbedding(head_size, rotary_dim,
max_position, base,
is_neox_style,
scaling_factor, dtype)
elif scaling_type == "dynamic":
scaling_factor = rope_scaling["factor"]
rotary_emb = DynamicNTKScalingRotaryEmbedding(
head_size, rotary_dim, max_position, base, is_neox_style,
scaling_factor, dtype)
elif scaling_type == "yarn":
scaling_factor = rope_scaling["factor"]
original_max_position = rope_scaling[
"original_max_position_embeddings"]
extra_kwargs = {
@@ -961,6 +981,7 @@ def get_rope(
scaling_factor, dtype,
**extra_kwargs)
elif scaling_type == "deepseek_yarn":
scaling_factor = rope_scaling["factor"]
original_max_position = rope_scaling[
"original_max_position_embeddings"]
# assert max_position == original_max_position * scaling_factor
@@ -973,9 +994,7 @@ def get_rope(
rotary_emb = DeepseekScalingRotaryEmbedding(
head_size, rotary_dim, original_max_position, base,
is_neox_style, scaling_factor, dtype, **extra_kwargs)
# The correct one should be "longrope" but keep "su" here
# for backward compatible
elif scaling_type == "su" or scaling_type == "longrope":
elif scaling_type == "longrope":
short_factor = rope_scaling["short_factor"]
long_factor = rope_scaling["long_factor"]
original_max_position = rope_scaling[
@@ -989,16 +1008,6 @@ def get_rope(
head_size, rotary_dim, max_position, original_max_position,
base, is_neox_style, dtype, short_factor, long_factor,
**extra_kwargs)
elif scaling_type == "mrope":
rotary_emb = MRotaryEmbedding(
head_size,
rotary_dim,
max_position,
base,
is_neox_style,
dtype,
mrope_section=rope_scaling["mrope_section"],
)
else:
raise ValueError(f"Unknown RoPE scaling type {scaling_type}")
_ROPE_DICT[key] = rotary_emb
+1 -1
View File
@@ -242,7 +242,7 @@ class DeepseekV2Attention(nn.Module):
bias=False,
quant_config=quant_config,
prefix=f"{prefix}.o_proj")
rope_scaling['type'] = 'deepseek_yarn'
rope_scaling["rope_type"] = 'deepseek_yarn'
self.rotary_emb = get_rope(qk_rope_head_dim,
rotary_dim=qk_rope_head_dim,
max_position=max_position_embeddings,
+1 -1
View File
@@ -179,7 +179,7 @@ class Phi3SmallSelfAttention(nn.Module):
rope_scaling["factor"] = self.rope_position_scale
else:
rope_scaling = {
"type": "linear",
"rope_type": "linear",
"factor": self.rope_position_scale,
}
+4 -4
View File
@@ -34,6 +34,8 @@ from PIL import Image
from transformers.image_utils import (get_image_size,
infer_channel_dimension_format,
to_numpy_array)
from transformers.models.qwen2_vl.configuration_qwen2_vl import (
Qwen2VLConfig, Qwen2VLVisionConfig)
from transformers.models.qwen2_vl.image_processing_qwen2_vl import (
make_batched_images, make_batched_videos, smart_resize)
@@ -62,8 +64,7 @@ from vllm.multimodal.base import MultiModalData
from vllm.multimodal.image import cached_get_image_processor
from vllm.platforms import current_platform
from vllm.sequence import IntermediateTensors, SequenceData
from vllm.transformers_utils.configs.qwen2vl import (Qwen2VLConfig,
Qwen2VLVisionConfig)
from vllm.transformers_utils.config import uses_mrope
from vllm.transformers_utils.processor import get_processor
from vllm.utils import is_cpu
@@ -1061,8 +1062,7 @@ class Qwen2VLForConditionalGeneration(nn.Module, SupportsMultiModal,
if image_input is None and video_input is None:
inputs_embeds = None
else:
rope_scaling = getattr(self.config, "rope_scaling", {})
if rope_scaling.get("type", None) == "mrope":
if uses_mrope(self.config):
assert positions.ndim == 2 and positions.size(0) == 3, (
"multimodal section rotary embedding requires "
f"(3, seq_len) positions, but got {positions.size()}")