Upload modular_isaac.py
Browse filesHello and congrats for the release!
This PR makes this model load with no additional dependency, you can also add a small inference notebook I made by replacing the username to Perceptron: https://colab.research.google.com/drive/1BHl2ZT8cYZ0HlP_q4HllFuCXWIBX_R_2?usp=sharing
if you add the "notebook.ipynb" repo to it's one-click open in the repository, making it easier for people to try out your model as well!
- modular_isaac.py +943 -21
modular_isaac.py
CHANGED
|
@@ -1,7 +1,7 @@
|
|
| 1 |
from __future__ import annotations
|
| 2 |
|
| 3 |
from collections import defaultdict
|
| 4 |
-
from typing import Any, Union, TypedDict
|
| 5 |
|
| 6 |
import math
|
| 7 |
import numpy as np
|
|
@@ -33,22 +33,944 @@ from transformers.models.siglip2.modeling_siglip2 import (
|
|
| 33 |
Siglip2MLP,
|
| 34 |
)
|
| 35 |
from transformers.models.siglip2.configuration_siglip2 import Siglip2VisionConfig
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 52 |
|
| 53 |
|
| 54 |
class PixelShuffleSiglip2VisionConfig(Siglip2VisionConfig):
|
|
@@ -474,7 +1396,7 @@ class Siglip2SequenceVisionTransformer(nn.Module):
|
|
| 474 |
# Configuration
|
| 475 |
# ============================================================================
|
| 476 |
|
| 477 |
-
MAX_PIXELS = 60_000_000 # 60
|
| 478 |
|
| 479 |
# Vision preprocessing constants
|
| 480 |
VISION_MEAN = (0.5, 0.5, 0.5)
|
|
@@ -491,13 +1413,13 @@ def _make_writeable(arr: np.ndarray) -> np.ndarray:
|
|
| 491 |
if arr.flags.writeable:
|
| 492 |
return arr
|
| 493 |
|
| 494 |
-
# First, try the cheap path — in
|
| 495 |
# and some shared memory buffers):
|
| 496 |
try:
|
| 497 |
arr.setflags(write=True)
|
| 498 |
return arr # success: no data copy
|
| 499 |
except ValueError:
|
| 500 |
-
# Buffer is inherently read
|
| 501 |
return arr.copy()
|
| 502 |
|
| 503 |
|
|
@@ -1623,4 +2545,4 @@ __all__ = [
|
|
| 1623 |
"IsaacModel",
|
| 1624 |
"IsaacForConditionalGeneration",
|
| 1625 |
"IsaacProcessor",
|
| 1626 |
-
]
|
|
|
|
| 1 |
from __future__ import annotations
|
| 2 |
|
| 3 |
from collections import defaultdict
|
| 4 |
+
from typing import Any, NewType, Union, TypedDict
|
| 5 |
|
| 6 |
import math
|
| 7 |
import numpy as np
|
|
|
|
| 33 |
Siglip2MLP,
|
| 34 |
)
|
| 35 |
from transformers.models.siglip2.configuration_siglip2 import Siglip2VisionConfig
|
| 36 |
+
|
| 37 |
+
import itertools
|
| 38 |
+
from collections.abc import Callable, Iterable
|
| 39 |
+
|
| 40 |
+
|
| 41 |
+
import heapq
|
| 42 |
+
from collections.abc import Callable, Iterable
|
| 43 |
+
from dataclasses import dataclass, field, fields, replace
|
| 44 |
+
from enum import Enum
|
| 45 |
+
|
| 46 |
+
from torch.profiler import record_function
|
| 47 |
+
|
| 48 |
+
|
| 49 |
+
class ModalityType(Enum):
|
| 50 |
+
"""
|
| 51 |
+
Base class for modality-type enumerations.
|
| 52 |
+
Each derived class (VisionType, TextType) holds
|
| 53 |
+
an integer value that identifies a specific modality.
|
| 54 |
+
Example usage:
|
| 55 |
+
If you have an object `my_event` of class `Event`,
|
| 56 |
+
you might write:
|
| 57 |
+
if my_event.type == VisionType.image:
|
| 58 |
+
# process an image frame
|
| 59 |
+
The methods below implement ordering and hashing
|
| 60 |
+
based on the integer `.value` of each enum member.
|
| 61 |
+
"""
|
| 62 |
+
|
| 63 |
+
@property
|
| 64 |
+
def modality(self):
|
| 65 |
+
return self.__class__
|
| 66 |
+
|
| 67 |
+
def __lt__(self, other):
|
| 68 |
+
if isinstance(other, ModalityType):
|
| 69 |
+
return self.value < other.value
|
| 70 |
+
raise NotImplementedError()
|
| 71 |
+
|
| 72 |
+
def __eq__(self, other):
|
| 73 |
+
if isinstance(other, ModalityType):
|
| 74 |
+
return self.value == other.value
|
| 75 |
+
raise NotImplementedError()
|
| 76 |
+
|
| 77 |
+
def __hash__(self):
|
| 78 |
+
return hash(self.value)
|
| 79 |
+
|
| 80 |
+
|
| 81 |
+
# NOTE: modality types need to be unique
|
| 82 |
+
class VisionType(ModalityType):
|
| 83 |
+
"""
|
| 84 |
+
Enum for vision modalities such as key video frames.
|
| 85 |
+
Typically used in video processing or image sequences.
|
| 86 |
+
Members:
|
| 87 |
+
image: A single image frame.
|
| 88 |
+
"""
|
| 89 |
+
|
| 90 |
+
image = 0
|
| 91 |
+
|
| 92 |
+
|
| 93 |
+
class TextType(ModalityType):
|
| 94 |
+
"""
|
| 95 |
+
Enum for text tokens and padding.
|
| 96 |
+
Members:
|
| 97 |
+
text: Actual textual tokens.
|
| 98 |
+
padding: Padding tokens used in sequence batching.
|
| 99 |
+
"""
|
| 100 |
+
|
| 101 |
+
text = 1
|
| 102 |
+
padding = 2
|
| 103 |
+
|
| 104 |
+
|
| 105 |
+
# maps idx -> type
|
| 106 |
+
ALL_TYPES = [
|
| 107 |
+
tp
|
| 108 |
+
for types in [
|
| 109 |
+
list(VisionType),
|
| 110 |
+
list(TextType),
|
| 111 |
+
]
|
| 112 |
+
for tp in types
|
| 113 |
+
]
|
| 114 |
+
|
| 115 |
+
|
| 116 |
+
# @dataclass
|
| 117 |
+
@dataclass(slots=True)
|
| 118 |
+
class Event:
|
| 119 |
+
"""
|
| 120 |
+
Represents a single data occurrence (with a specific type, time interval, and data payload).
|
| 121 |
+
Attributes:
|
| 122 |
+
data (Any): The actual data payload (e.g. a torch.Tensor, a string, etc.).
|
| 123 |
+
type (ModalityType): The modality type of the data (e.g., VisionType.image).
|
| 124 |
+
time (Tuple[float, float]): (start_time, end_time) indicating when this Event occurs.
|
| 125 |
+
role (Optional[str]): The role associated with this event (e.g., "user", "agent", "system").
|
| 126 |
+
If None, the event is always included in loss calculation.
|
| 127 |
+
Example usage:
|
| 128 |
+
evt = Event(data=torch.zeros((1, 224, 224, 3)), # e.g. a single image frame
|
| 129 |
+
type=VisionType.image,
|
| 130 |
+
time=(0.0, 0.04),
|
| 131 |
+
role="user")
|
| 132 |
+
"""
|
| 133 |
+
|
| 134 |
+
# Descriptors
|
| 135 |
+
data: Any
|
| 136 |
+
time: tuple[float, float]
|
| 137 |
+
type: ModalityType
|
| 138 |
+
role: str | None = None
|
| 139 |
+
|
| 140 |
+
# Structure
|
| 141 |
+
dims_virtual: list[int] | None = None # virtual/processed dimensions (e.g., pixel-shuffled)
|
| 142 |
+
dims_real: list[int] | None = None # real/actual tensor dimensions
|
| 143 |
+
idx_range: tuple[int, int] | None = None
|
| 144 |
+
|
| 145 |
+
# Misc Tags (data source, shard idx, etc.)
|
| 146 |
+
tags: dict = field(default_factory=dict)
|
| 147 |
+
|
| 148 |
+
def dims(self, virtual: bool = True) -> list[int] | None:
|
| 149 |
+
"""
|
| 150 |
+
Get the dimensions of this event.
|
| 151 |
+
Args:
|
| 152 |
+
virtual: If True (default), return virtual/processed dimensions (e.g., pixel-shuffled).
|
| 153 |
+
If False, return real/actual tensor dimensions.
|
| 154 |
+
Returns:
|
| 155 |
+
Dimensions list or None if not measured.
|
| 156 |
+
"""
|
| 157 |
+
if virtual:
|
| 158 |
+
return self.dims_virtual
|
| 159 |
+
else:
|
| 160 |
+
return self.dims_real
|
| 161 |
+
|
| 162 |
+
@property
|
| 163 |
+
def is_measured(self):
|
| 164 |
+
return self.dims_virtual is not None
|
| 165 |
+
|
| 166 |
+
def slice_tokens(self, start: int | None = None, end: int | None = None):
|
| 167 |
+
"""
|
| 168 |
+
Converts into a partial event where the only valid data is between start and end indices of the flattened data
|
| 169 |
+
"""
|
| 170 |
+
assert self.is_measured
|
| 171 |
+
assert start is not None and end is not None
|
| 172 |
+
assert self.idx_range[0] <= start <= end <= self.idx_range[1]
|
| 173 |
+
self.idx_range = (start or 0, end or math.prod(self.dims()))
|
| 174 |
+
|
| 175 |
+
def num_tokens(self, partial=True, virtual=True) -> int:
|
| 176 |
+
if not virtual:
|
| 177 |
+
assert partial is False and isinstance(self.data, torch.Tensor)
|
| 178 |
+
return math.prod(self.dims(virtual=False))
|
| 179 |
+
return self.idx_range[1] - self.idx_range[0] if partial else math.prod(self.dims())
|
| 180 |
+
|
| 181 |
+
def shallow_copy(self) -> Event:
|
| 182 |
+
return replace(self)
|
| 183 |
+
|
| 184 |
+
def __hash__(self) -> int:
|
| 185 |
+
"""Hash Event based on structure, excluding data."""
|
| 186 |
+
|
| 187 |
+
def make_hashable(obj):
|
| 188 |
+
"""Convert any object to hashable form."""
|
| 189 |
+
if obj is None:
|
| 190 |
+
return None
|
| 191 |
+
elif isinstance(obj, str | int | float | bool | tuple):
|
| 192 |
+
return obj
|
| 193 |
+
elif isinstance(obj, list):
|
| 194 |
+
return tuple(make_hashable(item) for item in obj) if obj else None
|
| 195 |
+
elif isinstance(obj, dict):
|
| 196 |
+
return tuple(sorted((k, make_hashable(v)) for k, v in obj.items())) if obj else None
|
| 197 |
+
elif hasattr(obj, "value"): # Enum types
|
| 198 |
+
return obj.value
|
| 199 |
+
else:
|
| 200 |
+
return str(obj) # Fallback for other types
|
| 201 |
+
|
| 202 |
+
hash_values = []
|
| 203 |
+
for fld in fields(self):
|
| 204 |
+
if fld.name == "data":
|
| 205 |
+
continue # Skip tensor data
|
| 206 |
+
|
| 207 |
+
value = getattr(self, fld.name)
|
| 208 |
+
hash_values.append(make_hashable(value))
|
| 209 |
+
|
| 210 |
+
return hash(tuple(hash_values))
|
| 211 |
+
|
| 212 |
+
def __eq__(self, other) -> bool:
|
| 213 |
+
"""
|
| 214 |
+
Compares two Event objects for strict equality,
|
| 215 |
+
allowing for float tolerances in torch.Tensors (via torch.allclose).
|
| 216 |
+
"""
|
| 217 |
+
if not isinstance(other, Event):
|
| 218 |
+
return False
|
| 219 |
+
|
| 220 |
+
for fld in fields(self):
|
| 221 |
+
self_value = getattr(self, fld.name)
|
| 222 |
+
other_value = getattr(other, fld.name)
|
| 223 |
+
|
| 224 |
+
if fld.name == "data":
|
| 225 |
+
# Special handling for tensor data with float tolerance
|
| 226 |
+
if isinstance(self_value, torch.Tensor) and isinstance(other_value, torch.Tensor):
|
| 227 |
+
if not torch.allclose(self_value, other_value):
|
| 228 |
+
return False
|
| 229 |
+
else:
|
| 230 |
+
if self_value != other_value:
|
| 231 |
+
return False
|
| 232 |
+
elif fld.name == "role":
|
| 233 |
+
# Special handling for role: both must be None or both must be set and equal
|
| 234 |
+
if (self_value is None) != (other_value is None):
|
| 235 |
+
return False
|
| 236 |
+
if self_value is not None and self_value != other_value:
|
| 237 |
+
return False
|
| 238 |
+
else:
|
| 239 |
+
# Standard equality for all other fields
|
| 240 |
+
if self_value != other_value:
|
| 241 |
+
return False
|
| 242 |
+
|
| 243 |
+
return True
|
| 244 |
+
|
| 245 |
+
|
| 246 |
+
@dataclass
|
| 247 |
+
class Stream:
|
| 248 |
+
"""
|
| 249 |
+
Represents an ordered sequence of Event objects, each with
|
| 250 |
+
a specific ModalityType and a time range.
|
| 251 |
+
Attributes:
|
| 252 |
+
events (List[Event]): The list of Event objects in the stream.
|
| 253 |
+
priority (List[ModalityType]): A list of modality types that define
|
| 254 |
+
how we might want to reorder or prioritize events if scheduling is needed.
|
| 255 |
+
Example usage:
|
| 256 |
+
# Create two events of different types
|
| 257 |
+
evt1 = Event(torch.zeros((1, 224, 224, 3)), VisionType.image, (0.0, 0.04))
|
| 258 |
+
evt2 = Event(torch.randint(0, 1000, (16, 1)), TextType.text, (0.0, 0.32))
|
| 259 |
+
# Make a stream with a given priority
|
| 260 |
+
s = Stream(events=[evt1, evt2],
|
| 261 |
+
priority=[VisionType.image, TextType.text])
|
| 262 |
+
print(s)
|
| 263 |
+
"""
|
| 264 |
+
|
| 265 |
+
events: list[Event]
|
| 266 |
+
priority: list[ModalityType] # priority of stream ordering
|
| 267 |
+
|
| 268 |
+
def __len__(self):
|
| 269 |
+
"""Returns the number of Event objects in this Stream."""
|
| 270 |
+
return len(self.events)
|
| 271 |
+
|
| 272 |
+
def __getitem__(self, key: int) -> Stream | Event:
|
| 273 |
+
return self.events[key]
|
| 274 |
+
|
| 275 |
+
def __iter__(self):
|
| 276 |
+
"""
|
| 277 |
+
Yields each Event in the Stream, enabling iteration like:
|
| 278 |
+
for event in my_stream:
|
| 279 |
+
...
|
| 280 |
+
"""
|
| 281 |
+
yield from self.events
|
| 282 |
+
|
| 283 |
+
# --- after ------------------------------------------------------------
|
| 284 |
+
@record_function("Stream.map")
|
| 285 |
+
def map(
|
| 286 |
+
self,
|
| 287 |
+
func: Callable[[Event], dict[str, Any]],
|
| 288 |
+
*,
|
| 289 |
+
copy_unchanged: bool = False, # opt-in if you really need isolation
|
| 290 |
+
) -> Stream:
|
| 291 |
+
"""
|
| 292 |
+
Apply *func* to every event and return a new Stream.
|
| 293 |
+
*func* must return a **dict of fields that actually change**.
|
| 294 |
+
We create **one shallow copy** only when something changes;
|
| 295 |
+
unchanged events are reused directly, which is inexpensive and
|
| 296 |
+
keeps autograd graphs intact.
|
| 297 |
+
"""
|
| 298 |
+
mapped: list[Event] = []
|
| 299 |
+
for ev in self.events:
|
| 300 |
+
delta = func(ev)
|
| 301 |
+
if not delta: # fast-path: nothing changes
|
| 302 |
+
mapped.append(ev if not copy_unchanged else ev.shallow_copy())
|
| 303 |
+
continue
|
| 304 |
+
|
| 305 |
+
new_ev = ev.shallow_copy() # ⚡ no tensor clone
|
| 306 |
+
for k, v in delta.items():
|
| 307 |
+
setattr(new_ev, k, v)
|
| 308 |
+
mapped.append(new_ev)
|
| 309 |
+
|
| 310 |
+
return create_stream(mapped, priority=self.priority, schedule=False)
|
| 311 |
+
|
| 312 |
+
@record_function("Stream.compact")
|
| 313 |
+
def compact(self) -> torch.Tensor:
|
| 314 |
+
assert all([(isinstance(ev.data, torch.Tensor) and ev.is_measured) for ev in self.events]), (
|
| 315 |
+
"Stream.compact only works for streams with events that have measured tensor data"
|
| 316 |
+
)
|
| 317 |
+
return torch.cat([ev.data for ev in self.events]).contiguous()
|
| 318 |
+
|
| 319 |
+
@record_function("Stream.map_compact")
|
| 320 |
+
def map_compact(self, event_tf: Callable[[Event], list[Any]]) -> torch.Tensor:
|
| 321 |
+
mapped_list = []
|
| 322 |
+
for event in self:
|
| 323 |
+
mapped_list.extend(event_tf(event))
|
| 324 |
+
tensor = torch.tensor(
|
| 325 |
+
mapped_list,
|
| 326 |
+
dtype=torch.long,
|
| 327 |
+
device=next(
|
| 328 |
+
(ev.data.device for ev in self.events if isinstance(ev.data, torch.Tensor)),
|
| 329 |
+
"cpu",
|
| 330 |
+
),
|
| 331 |
+
).contiguous()
|
| 332 |
+
return tensor
|
| 333 |
+
|
| 334 |
+
def flatten(self) -> Stream:
|
| 335 |
+
return self.map(lambda ev: {"data": ev.data.reshape(-1, ev.data.shape[-1])})
|
| 336 |
+
|
| 337 |
+
def shallow_copy(self) -> Stream:
|
| 338 |
+
events_copy = [ev.shallow_copy() for ev in self.events]
|
| 339 |
+
return create_stream(events=events_copy, priority=self.priority, schedule=False)
|
| 340 |
+
|
| 341 |
+
def __hash__(self) -> int:
|
| 342 |
+
"""Hash Stream based on structure."""
|
| 343 |
+
return hash(
|
| 344 |
+
(
|
| 345 |
+
tuple(p.value for p in self.priority), # Convert enums to values
|
| 346 |
+
tuple(hash(event) for event in self.events), # Use Event.__hash__
|
| 347 |
+
)
|
| 348 |
+
)
|
| 349 |
+
|
| 350 |
+
def __eq__(self, other) -> bool:
|
| 351 |
+
"""Compare Streams structurally."""
|
| 352 |
+
if not isinstance(other, Stream):
|
| 353 |
+
return False
|
| 354 |
+
|
| 355 |
+
return (
|
| 356 |
+
self.priority == other.priority
|
| 357 |
+
and len(self.events) == len(other.events)
|
| 358 |
+
and all(e1 == e2 for e1, e2 in zip(self.events, other.events, strict=False))
|
| 359 |
+
)
|
| 360 |
+
|
| 361 |
+
|
| 362 |
+
# TODO: implement all types of cool indexing which can happen since TensorStream assuems Event.data = Tensor
|
| 363 |
+
@dataclass
|
| 364 |
+
class TensorStream:
|
| 365 |
+
streams: list[Stream]
|
| 366 |
+
_device: torch.device | None = None
|
| 367 |
+
|
| 368 |
+
def __post_init__(self):
|
| 369 |
+
for stream in self.streams:
|
| 370 |
+
for event in stream.events:
|
| 371 |
+
assert isinstance(event.data, torch.Tensor)
|
| 372 |
+
if self._device is None:
|
| 373 |
+
self._device = torch.device(event.data.device)
|
| 374 |
+
|
| 375 |
+
# TODO: implement non-strict compaction modes
|
| 376 |
+
@record_function("TensorStream.compact")
|
| 377 |
+
def compact(self, mode="strict") -> torch.Tensor:
|
| 378 |
+
compact_tensor_stream = torch.stack([stream.compact() for stream in self.streams]).contiguous()
|
| 379 |
+
return compact_tensor_stream
|
| 380 |
+
|
| 381 |
+
@record_function("TensorStream.map")
|
| 382 |
+
def map(self, event_tf: Callable[[Event], dict[str, Any]]) -> TensorStream:
|
| 383 |
+
mapped_streams = [stream.map(event_tf) for stream in self.streams]
|
| 384 |
+
return TensorStream(mapped_streams)
|
| 385 |
+
|
| 386 |
+
@record_function("TensorStream.map_compact")
|
| 387 |
+
def map_compact(self, event_tf: Callable[[Event], list[Any]]) -> torch.Tensor:
|
| 388 |
+
mapped_list = []
|
| 389 |
+
for stream in self.streams:
|
| 390 |
+
for event in stream:
|
| 391 |
+
mapped_list.extend(event_tf(event))
|
| 392 |
+
B, T = self.shape
|
| 393 |
+
tensor = torch.tensor(mapped_list, dtype=torch.long, device=self.device).reshape(B, T)
|
| 394 |
+
return tensor
|
| 395 |
+
|
| 396 |
+
def flat_stream(self) -> Stream:
|
| 397 |
+
if not self.streams:
|
| 398 |
+
return create_stream([], priority=[], schedule=False)
|
| 399 |
+
return create_stream(
|
| 400 |
+
[event for stream in self.streams for event in stream], priority=self.streams[0].priority, schedule=False
|
| 401 |
+
)
|
| 402 |
+
|
| 403 |
+
@property
|
| 404 |
+
def device(self):
|
| 405 |
+
return self._device
|
| 406 |
+
|
| 407 |
+
@property
|
| 408 |
+
def shape(self):
|
| 409 |
+
seq_lens = [sum([ev.num_tokens() for ev in stream]) for stream in self.streams]
|
| 410 |
+
assert all([sl == seq_lens[0] for sl in seq_lens]), (
|
| 411 |
+
f"each stream must have same token count to have a shape: {seq_lens}"
|
| 412 |
+
)
|
| 413 |
+
return (len(seq_lens), seq_lens[0])
|
| 414 |
+
|
| 415 |
+
@record_function("TensorStream.to")
|
| 416 |
+
def to(
|
| 417 |
+
self,
|
| 418 |
+
device: torch.device | str,
|
| 419 |
+
dtype: torch.dtype | None = None,
|
| 420 |
+
non_blocking: bool = True,
|
| 421 |
+
) -> TensorStream:
|
| 422 |
+
"""
|
| 423 |
+
Move **all** `Event.data` tensors to *device*.
|
| 424 |
+
We send each tensor individually instead of the
|
| 425 |
+
flatten → unflatten round-trip:
|
| 426 |
+
* one async H2D copy per tensor (still overlapped when
|
| 427 |
+
`pin_memory=True` is set on the DataLoader),
|
| 428 |
+
* no extra host-side concat, no extra device allocation,
|
| 429 |
+
* `requires_grad` flags are preserved.
|
| 430 |
+
NOTE: textual modalities are always cast to `torch.long`;
|
| 431 |
+
everything else keeps its original
|
| 432 |
+
dtype unless an explicit *dtype* argument is supplied.
|
| 433 |
+
"""
|
| 434 |
+
target_device = torch.device(device)
|
| 435 |
+
|
| 436 |
+
for stream in self.streams:
|
| 437 |
+
for ev in stream:
|
| 438 |
+
# ------------------------------------------------------------------
|
| 439 |
+
# Decide the dtype for *this* event.
|
| 440 |
+
# ------------------------------------------------------------------
|
| 441 |
+
if ev.type in list(TextType):
|
| 442 |
+
tgt_dtype = torch.long
|
| 443 |
+
else:
|
| 444 |
+
tgt_dtype = dtype or ev.data.dtype
|
| 445 |
+
|
| 446 |
+
# ------------------------------------------------------------------
|
| 447 |
+
# Perform the device / dtype move.
|
| 448 |
+
# ------------------------------------------------------------------
|
| 449 |
+
# We clone no tensor here; torch will reuse storage
|
| 450 |
+
# if `dtype` and `device` are unchanged.
|
| 451 |
+
moved = ev.data.to(
|
| 452 |
+
device=target_device,
|
| 453 |
+
dtype=tgt_dtype,
|
| 454 |
+
non_blocking=non_blocking,
|
| 455 |
+
)
|
| 456 |
+
|
| 457 |
+
# Preserve autograd leaf & grad-enabled state.
|
| 458 |
+
moved.requires_grad_(ev.data.requires_grad)
|
| 459 |
+
|
| 460 |
+
ev.data = moved
|
| 461 |
+
|
| 462 |
+
# Remember where the whole TensorStream lives now.
|
| 463 |
+
self._device = target_device
|
| 464 |
+
return self
|
| 465 |
+
|
| 466 |
+
@record_function("TensorStream.pin_memory")
|
| 467 |
+
def pin_memory(self, non_blocking: bool = True) -> TensorStream:
|
| 468 |
+
"""
|
| 469 |
+
Page-lock (aka *pin*) all **CPU** tensors contained in this
|
| 470 |
+
`TensorStream`. Pinned tensors make subsequent asynchronous
|
| 471 |
+
H2D copies (e.g. inside `TensorStream.to("cuda")`) faster and,
|
| 472 |
+
when used together with a `DataLoader(pin_memory=True)`,
|
| 473 |
+
enable overlap of host-to-device transfers with GPU execution.
|
| 474 |
+
The call is a no-op for tensors that are already on a CUDA /
|
| 475 |
+
MPS / other non-CPU device.
|
| 476 |
+
Parameters
|
| 477 |
+
----------
|
| 478 |
+
non_blocking : bool, default = True
|
| 479 |
+
Forwarded to `Tensor.pin_memory()`; should almost always
|
| 480 |
+
stay *True* so later `to(device, non_blocking=True)` calls
|
| 481 |
+
can overlap.
|
| 482 |
+
Returns
|
| 483 |
+
-------
|
| 484 |
+
self : TensorStream
|
| 485 |
+
The same object (mutated in-place) to allow call chaining.
|
| 486 |
+
"""
|
| 487 |
+
for stream in self.streams:
|
| 488 |
+
for ev in stream:
|
| 489 |
+
if ev.data.device.type == "cpu":
|
| 490 |
+
# `pin_memory()` clones only when needed
|
| 491 |
+
pinned = ev.data.pin_memory() # noqa: F841
|
| 492 |
+
# NB: pin_memory() preserves dtype/shape/grad/etc.
|
| 493 |
+
if not non_blocking:
|
| 494 |
+
# ensure the pinning work is done now
|
| 495 |
+
torch.cuda.current_stream().synchronize() # safe on CPU too
|
| 496 |
+
ev.data = pinned
|
| 497 |
+
# `_device` **stays** the same (still CPU) – no change needed
|
| 498 |
+
return self
|
| 499 |
+
|
| 500 |
+
def __hash__(self) -> int:
|
| 501 |
+
"""Hash TensorStream based on structure."""
|
| 502 |
+
return hash(
|
| 503 |
+
(
|
| 504 |
+
tuple(hash(stream) for stream in self.streams), # Use Stream.__hash__
|
| 505 |
+
str(self._device) if self._device else None,
|
| 506 |
+
self.shape,
|
| 507 |
+
)
|
| 508 |
+
)
|
| 509 |
+
|
| 510 |
+
def __eq__(self, other) -> bool:
|
| 511 |
+
"""Compare TensorStreams structurally."""
|
| 512 |
+
if not isinstance(other, TensorStream):
|
| 513 |
+
return False
|
| 514 |
+
|
| 515 |
+
return (
|
| 516 |
+
self._device == other._device
|
| 517 |
+
and self.shape == other.shape
|
| 518 |
+
and len(self.streams) == len(other.streams)
|
| 519 |
+
and all(s1 == s2 for s1, s2 in zip(self.streams, other.streams, strict=False))
|
| 520 |
+
)
|
| 521 |
+
|
| 522 |
+
|
| 523 |
+
def collate_tensor_stream(
|
| 524 |
+
tensor_streams: list[TensorStream],
|
| 525 |
+
) -> TensorStream:
|
| 526 |
+
return TensorStream([stream for ts in tensor_streams for stream in ts.streams])
|
| 527 |
+
|
| 528 |
+
|
| 529 |
+
def _schedule_stream(stream: Stream) -> Stream:
|
| 530 |
+
"""
|
| 531 |
+
Internal function that reorders (schedules) the events in a Stream
|
| 532 |
+
based on the stream's priority.
|
| 533 |
+
By default, this calls schedule_events(...) and reorders the events accordingly.
|
| 534 |
+
The new ordering is assigned in-place to stream.events.
|
| 535 |
+
Example usage (indirect):
|
| 536 |
+
new_stream = _schedule_stream(old_stream)
|
| 537 |
+
"""
|
| 538 |
+
scheduled_inds = schedule_events(stream, priority=stream.priority)
|
| 539 |
+
stream.events = [stream.events[i] for i in scheduled_inds]
|
| 540 |
+
return stream
|
| 541 |
+
|
| 542 |
+
|
| 543 |
+
def create_stream(events: list[Event], priority: list[ModalityType], schedule: bool = True) -> Stream:
|
| 544 |
+
"""
|
| 545 |
+
Creates a new Stream with the given events and priority.
|
| 546 |
+
If 'schedule' is True, the events are reordered by calling _schedule_stream.
|
| 547 |
+
Example usage:
|
| 548 |
+
evt1 = Event(torch.zeros(10), TextType.text, (0.0, 1.0))
|
| 549 |
+
evt2 = Event(torch.ones(10), TextType.text, (1.0, 2.0))
|
| 550 |
+
my_stream = create_stream(events=[evt1, evt2],
|
| 551 |
+
priority=[TextType.text],
|
| 552 |
+
schedule=False)
|
| 553 |
+
print(my_stream)
|
| 554 |
+
"""
|
| 555 |
+
stream = Stream(events, priority)
|
| 556 |
+
if schedule:
|
| 557 |
+
stream = _schedule_stream(stream)
|
| 558 |
+
return stream
|
| 559 |
+
|
| 560 |
+
|
| 561 |
+
def merge_streams(streams: Iterable[Stream]) -> Stream:
|
| 562 |
+
"""
|
| 563 |
+
Merges multiple Stream objects into one.
|
| 564 |
+
The priority of the merged stream is chosen from the longest priority list among the inputs.
|
| 565 |
+
Stream priorities must be consistent with the chosen priority.
|
| 566 |
+
All events are concatenated, and a new Stream is created (and scheduled).
|
| 567 |
+
Example usage:
|
| 568 |
+
merged = merge_streams([stream1, stream2])
|
| 569 |
+
"""
|
| 570 |
+
chosen_priority = max([stream.priority for stream in streams], key=len)
|
| 571 |
+
assert all(
|
| 572 |
+
[str(stream.priority) in str([p for p in chosen_priority if p in stream.priority]) for stream in streams]
|
| 573 |
+
), "One or more streams has a priority order that doesn't match the merged stream"
|
| 574 |
+
merged_event_list = [ev for stream in streams for ev in stream.events]
|
| 575 |
+
merged_stream = create_stream(merged_event_list, chosen_priority) # non-root stream creation
|
| 576 |
+
return merged_stream
|
| 577 |
+
|
| 578 |
+
|
| 579 |
+
EventDescriptor = NewType("EventDescriptor", Any)
|
| 580 |
+
|
| 581 |
+
|
| 582 |
+
# NOTE: actually not used now but thought it *might* be useful
|
| 583 |
+
def get_stream_descriptor(
|
| 584 |
+
stream: Stream, measure_fn: Callable[[Event], EventDescriptor] = lambda ev: ev.type
|
| 585 |
+
) -> set[Any]:
|
| 586 |
+
"""
|
| 587 |
+
Create a set of descriptors for each Event in a Stream based on measure_fn.
|
| 588 |
+
measure_fn maps an Event to a descriptive key.
|
| 589 |
+
For example, if events have different data shapes, one might use:
|
| 590 |
+
measure_fn = lambda ev: ev.data.shape
|
| 591 |
+
i.e.
|
| 592 |
+
stream of VisionTypes with tensors of shapes [(1, 3, 3), (1, 3, 3), (1, 4, 4)]
|
| 593 |
+
get_stream_descriptor(stream, measure_fn=lambda t: t.shape) = {(1, 3, 3), (1, 4, 4)}
|
| 594 |
+
now we can pass this into group_streams which will split out vision sub-streams which can be bundled
|
| 595 |
+
Returns:
|
| 596 |
+
A set of descriptors representing the Events in the stream.
|
| 597 |
+
Example usage:
|
| 598 |
+
descriptor = get_stream_descriptor(my_stream, lambda ev: ev.type)
|
| 599 |
+
"""
|
| 600 |
+
stream_descriptor = set()
|
| 601 |
+
for ev in stream.events:
|
| 602 |
+
ev_measurement = measure_fn(ev)
|
| 603 |
+
stream_descriptor.add(ev_measurement)
|
| 604 |
+
return stream_descriptor
|
| 605 |
+
|
| 606 |
+
|
| 607 |
+
def group_streams(
|
| 608 |
+
stream: Stream, group_fn: Callable[[Event], EventDescriptor], schedule=True
|
| 609 |
+
) -> dict[EventDescriptor, Stream]:
|
| 610 |
+
"""
|
| 611 |
+
Splits a single Stream into multiple sub-Streams, grouped by the output of group_fn(event).
|
| 612 |
+
For example, group_fn could be:
|
| 613 |
+
- lambda ev: ev.type
|
| 614 |
+
- lambda ev: ev.type.modality
|
| 615 |
+
- lambda ev: (ev.type.modality, ev.data.shape)
|
| 616 |
+
Returns:
|
| 617 |
+
A dictionary mapping each group key to a Stream of events belonging to that group.
|
| 618 |
+
If 'schedule' is True, each sub-Stream is scheduled via create_stream(..., schedule=True).
|
| 619 |
+
Example usage:
|
| 620 |
+
substreams = group_streams(my_stream, lambda ev: ev.type)
|
| 621 |
+
"""
|
| 622 |
+
split_streams: defaultdict[EventDescriptor, list[Event]] = defaultdict(list)
|
| 623 |
+
for ev in stream:
|
| 624 |
+
group = group_fn(ev)
|
| 625 |
+
split_streams[group].append(ev)
|
| 626 |
+
for g, events in split_streams.items():
|
| 627 |
+
split_streams[g] = create_stream(events, stream.priority, schedule=schedule)
|
| 628 |
+
return dict(split_streams)
|
| 629 |
+
|
| 630 |
+
|
| 631 |
+
# Define Category for clarity
|
| 632 |
+
Category = NewType("Category", Any)
|
| 633 |
+
|
| 634 |
+
|
| 635 |
+
def schedule_events(stream: Stream, priority: list[Category]) -> list[int]:
|
| 636 |
+
"""
|
| 637 |
+
Schedule events based on their start time and priority using a topological sort algorithm.
|
| 638 |
+
The priority list defines the ordering of categories.
|
| 639 |
+
This function:
|
| 640 |
+
1. Pairs each event with its original index.
|
| 641 |
+
2. Sorts events by start time.
|
| 642 |
+
3. Builds a dependency graph based on overlapping events.
|
| 643 |
+
4. Uses a heap to perform a deterministic topological sort with tie-breakers.
|
| 644 |
+
Raises:
|
| 645 |
+
ValueError: If a cycle is detected in the events (i.e., no valid ordering exists).
|
| 646 |
+
Returns:
|
| 647 |
+
List[int]: A list of original indices representing the scheduled order of events.
|
| 648 |
+
"""
|
| 649 |
+
priority_index: dict[Category, int] = {category: idx for idx, category in enumerate(priority)}
|
| 650 |
+
|
| 651 |
+
# Pair each event metadata with its original index
|
| 652 |
+
events = []
|
| 653 |
+
for i, event in enumerate(stream.events):
|
| 654 |
+
events.append(
|
| 655 |
+
(
|
| 656 |
+
i,
|
| 657 |
+
event.time[0],
|
| 658 |
+
event.time[1],
|
| 659 |
+
event.type,
|
| 660 |
+
)
|
| 661 |
+
)
|
| 662 |
+
|
| 663 |
+
sorted_events = sorted(events, key=lambda e: e[1]) # sort by start time
|
| 664 |
+
num_events = len(sorted_events)
|
| 665 |
+
|
| 666 |
+
# Build dependency graph
|
| 667 |
+
graph = defaultdict(set)
|
| 668 |
+
indegree = {i: 0 for i in range(num_events)}
|
| 669 |
+
|
| 670 |
+
for i in range(num_events):
|
| 671 |
+
idx_i, start_i, end_i, category_i = sorted_events[i]
|
| 672 |
+
prio_i = priority_index[category_i]
|
| 673 |
+
for j in range(i + 1, num_events):
|
| 674 |
+
idx_j, start_j, end_j, category_j = sorted_events[j]
|
| 675 |
+
if start_j >= end_i:
|
| 676 |
+
break
|
| 677 |
+
if end_i > start_j and end_j > start_i:
|
| 678 |
+
prio_j = priority_index[category_j]
|
| 679 |
+
if prio_i < prio_j:
|
| 680 |
+
graph[i].add(j)
|
| 681 |
+
indegree[j] += 1
|
| 682 |
+
elif prio_i > prio_j:
|
| 683 |
+
graph[j].add(i)
|
| 684 |
+
indegree[i] += 1
|
| 685 |
+
|
| 686 |
+
# Use heap for deterministic tie-breakers: (start_time, priority, original_index)
|
| 687 |
+
heap = [
|
| 688 |
+
(
|
| 689 |
+
sorted_events[i][1],
|
| 690 |
+
priority_index[sorted_events[i][3]],
|
| 691 |
+
sorted_events[i][0],
|
| 692 |
+
i,
|
| 693 |
+
)
|
| 694 |
+
for i in range(num_events)
|
| 695 |
+
if indegree[i] == 0
|
| 696 |
+
]
|
| 697 |
+
heapq.heapify(heap)
|
| 698 |
+
resolved_order = []
|
| 699 |
+
|
| 700 |
+
while heap:
|
| 701 |
+
_, _, _, u = heapq.heappop(heap)
|
| 702 |
+
resolved_order.append(u)
|
| 703 |
+
for v in graph[u]:
|
| 704 |
+
indegree[v] -= 1
|
| 705 |
+
if indegree[v] == 0:
|
| 706 |
+
heapq.heappush(
|
| 707 |
+
heap,
|
| 708 |
+
(
|
| 709 |
+
sorted_events[v][1],
|
| 710 |
+
priority_index[sorted_events[v][3]],
|
| 711 |
+
sorted_events[v][0],
|
| 712 |
+
v,
|
| 713 |
+
),
|
| 714 |
+
)
|
| 715 |
+
|
| 716 |
+
if len(resolved_order) != num_events:
|
| 717 |
+
raise ValueError("Cycle detected in events, cannot resolve order")
|
| 718 |
+
|
| 719 |
+
return [sorted_events[i][0] for i in resolved_order]
|
| 720 |
+
|
| 721 |
+
def compute_mrope_pos_tensor(ts: TensorStream, n_pos_dims: int = 3) -> torch.Tensor:
|
| 722 |
+
"""
|
| 723 |
+
Create a (batch, T, n_pos_dims) position tensor in one sweep.
|
| 724 |
+
The first dim is the running “time” index, the rest are spatial (or 1-fillers).
|
| 725 |
+
|
| 726 |
+
Args:
|
| 727 |
+
ts : TensorStream
|
| 728 |
+
n_pos_dims : total coordinate dimensions (default 3)
|
| 729 |
+
|
| 730 |
+
Returns:
|
| 731 |
+
torch.LongTensor - shape (batch_size, seq_len, n_pos_dims)
|
| 732 |
+
"""
|
| 733 |
+
|
| 734 |
+
# Manually iterate through streams and events like map_compact does,
|
| 735 |
+
# but maintain cumulative time offset for each stream
|
| 736 |
+
all_coords = []
|
| 737 |
+
for stream in ts.streams: # one Stream == one batch sample
|
| 738 |
+
cumulative_offset = 0 # running time index for this stream
|
| 739 |
+
|
| 740 |
+
for event in stream:
|
| 741 |
+
# --- build coordinate grid for THIS event using itertools (no tensor ops) ---
|
| 742 |
+
dims = (event.dims() or [1]) + [1] * (n_pos_dims - len(event.dims() or []))
|
| 743 |
+
|
| 744 |
+
# Create ranges for each dimension (similar to old _finalize implementation)
|
| 745 |
+
first_dim = range(cumulative_offset, cumulative_offset + dims[0])
|
| 746 |
+
cumulative_offset += dims[0] # advance time for the next event
|
| 747 |
+
other_dims = [range(d) for d in dims[1:]]
|
| 748 |
+
|
| 749 |
+
# Use itertools.product to create all coordinate combinations
|
| 750 |
+
full_coords = list(itertools.product(first_dim, *other_dims))
|
| 751 |
+
|
| 752 |
+
# Slice if the event is partial
|
| 753 |
+
s, e = event.idx_range
|
| 754 |
+
coords = full_coords[s:e]
|
| 755 |
+
|
| 756 |
+
# Extend the flattened coordinate list
|
| 757 |
+
all_coords.extend(coords)
|
| 758 |
+
|
| 759 |
+
# Convert to tensor and reshape to (B, T, n_pos_dims)
|
| 760 |
+
B, T = ts.shape
|
| 761 |
+
return torch.tensor(all_coords, dtype=torch.long, device=ts.device).reshape(B, T, n_pos_dims)
|
| 762 |
+
|
| 763 |
+
|
| 764 |
+
# ──────────────────────────────────────────────────────────────────────────
|
| 765 |
+
# Generic event-labelling helper
|
| 766 |
+
# ──────────────────────────────────────────────────────────────────────────
|
| 767 |
+
def event_mask(
|
| 768 |
+
ts: TensorStream,
|
| 769 |
+
tag_fn: Callable[[Event], int | None],
|
| 770 |
+
default: int = -1,
|
| 771 |
+
) -> torch.Tensor:
|
| 772 |
+
"""
|
| 773 |
+
Build a (batch, seq_len) LongTensor whose value for every *token*
|
| 774 |
+
is given by `tag_fn(event)`, falling back to `default` when the
|
| 775 |
+
function returns None.
|
| 776 |
+
|
| 777 |
+
The work is done in a single pass via `map → compact`.
|
| 778 |
+
"""
|
| 779 |
+
|
| 780 |
+
def to_label(ev: Event) -> Any:
|
| 781 |
+
label = tag_fn(ev)
|
| 782 |
+
if label is None:
|
| 783 |
+
label = default
|
| 784 |
+
return [label] * ev.num_tokens()
|
| 785 |
+
|
| 786 |
+
return ts.map_compact(to_label).squeeze(-1)
|
| 787 |
+
|
| 788 |
+
|
| 789 |
+
def event_mask_by_key(
|
| 790 |
+
ts: TensorStream,
|
| 791 |
+
key: str,
|
| 792 |
+
tag_index: dict[str, int],
|
| 793 |
+
default: int = -1,
|
| 794 |
+
) -> torch.Tensor:
|
| 795 |
+
"""
|
| 796 |
+
Faster call-site syntax when you just want to look up
|
| 797 |
+
`event.tags[key]` and map it through `tag_index`.
|
| 798 |
+
"""
|
| 799 |
+
return event_mask(
|
| 800 |
+
ts,
|
| 801 |
+
lambda ev: tag_index.get(ev.tags.get(key)) if key in ev.tags else None,
|
| 802 |
+
default=default,
|
| 803 |
+
)
|
| 804 |
+
|
| 805 |
+
|
| 806 |
+
def modality_mask(ts: TensorStream) -> torch.Tensor:
|
| 807 |
+
return event_mask(ts, lambda ev: ev.type.value)
|
| 808 |
+
|
| 809 |
+
|
| 810 |
+
ROLE_TO_IDX = {
|
| 811 |
+
None: -1,
|
| 812 |
+
"": -1,
|
| 813 |
+
"agent": 0,
|
| 814 |
+
"user": 1,
|
| 815 |
+
"system": 2,
|
| 816 |
+
# … add more if you like
|
| 817 |
+
}
|
| 818 |
+
|
| 819 |
+
|
| 820 |
+
def role_mask(ts: TensorStream) -> torch.Tensor:
|
| 821 |
+
return event_mask(ts, lambda ev: ROLE_TO_IDX.get(ev.role, -1))
|
| 822 |
+
|
| 823 |
+
|
| 824 |
+
def tensor_stream_token_view(ts: TensorStream) -> torch.Tensor:
|
| 825 |
+
"""
|
| 826 |
+
Return a (B, T) token view by summing across the last dim of every
|
| 827 |
+
event and flattening over the selected token range.
|
| 828 |
+
"""
|
| 829 |
+
|
| 830 |
+
def to_token_view(ev: Event) -> list[int]:
|
| 831 |
+
# collapse all but the last dim, cast to long
|
| 832 |
+
flat = ev.data.sum(dim=-1).long().reshape(-1)
|
| 833 |
+
if ev.idx_range is not None:
|
| 834 |
+
s, e = ev.idx_range
|
| 835 |
+
return flat[s:e].tolist()
|
| 836 |
+
else:
|
| 837 |
+
return flat.tolist()
|
| 838 |
+
|
| 839 |
+
return ts.map_compact(to_token_view) # shape (B, T)
|
| 840 |
+
|
| 841 |
+
|
| 842 |
+
def reconstruct_tensor_stream_from_compact_dict(
|
| 843 |
+
ts: TensorStream, compact_dict: dict[ModalityType, torch.Tensor]
|
| 844 |
+
) -> TensorStream:
|
| 845 |
+
streams = []
|
| 846 |
+
for stream in ts.streams:
|
| 847 |
+
event_list = []
|
| 848 |
+
for event in stream:
|
| 849 |
+
new_event = event.shallow_copy()
|
| 850 |
+
new_event.data = compact_dict[event.type][event.idx_range[0] : event.idx_range[1]]
|
| 851 |
+
compact_dict[event.type] = compact_dict[event.type][event.num_tokens(partial=False) :]
|
| 852 |
+
event_list.append(new_event)
|
| 853 |
+
streams.append(Stream(event_list, priority=stream.priority))
|
| 854 |
+
return TensorStream(streams)
|
| 855 |
+
|
| 856 |
+
|
| 857 |
+
def set_data(
|
| 858 |
+
tensor_stream: TensorStream,
|
| 859 |
+
stream_types: Iterable[ModalityType],
|
| 860 |
+
roles: Iterable[str] = ROLE_TO_IDX.keys(),
|
| 861 |
+
) -> tuple[torch.Tensor, torch.Tensor]:
|
| 862 |
+
"""
|
| 863 |
+
Gathers data from a TensorStream according to the given stream types
|
| 864 |
+
and returns (data, mask) where 'data' has valid entries for
|
| 865 |
+
each requested stream type and 'mask' indicates which elements
|
| 866 |
+
in 'data' are valid.
|
| 867 |
+
|
| 868 |
+
NOTE: Currently assumes stream_types are text-based types, but can be extended.
|
| 869 |
+
|
| 870 |
+
Args:
|
| 871 |
+
tensor_stream (TensorStream):
|
| 872 |
+
The input TensorStream which contains data for multiple modalities.
|
| 873 |
+
stream_types (Iterable[ModalityType]):
|
| 874 |
+
A list or iterable of modality types (e.g., TextType, VisionType, etc.)
|
| 875 |
+
to retrieve from the TensorStream.
|
| 876 |
+
exclude_non_agent_roles (bool, optional):
|
| 877 |
+
If True, only include tokens with role="agent" or role=None in the loss calculation.
|
| 878 |
+
Defaults to False.
|
| 879 |
+
|
| 880 |
+
Returns:
|
| 881 |
+
Tuple[torch.Tensor, torch.Tensor]:
|
| 882 |
+
- data: A tensor of the same shape as the internal metadata shape,
|
| 883 |
+
containing valid entries from the given stream types.
|
| 884 |
+
- mask: A boolean tensor of the same shape, where True indicates
|
| 885 |
+
the corresponding element in 'data' is valid/used.
|
| 886 |
+
"""
|
| 887 |
+
# Retrieve indexing and shape metadata
|
| 888 |
+
st_tensor = modality_mask(tensor_stream) # (B, T) modality-ids
|
| 889 |
+
roles_tensor = role_mask(tensor_stream) # (B, T) role-ids
|
| 890 |
+
|
| 891 |
+
# Create output data placeholders on the same device
|
| 892 |
+
data = torch.zeros_like(st_tensor).to(tensor_stream.device)
|
| 893 |
+
set_data_mask = torch.zeros_like(st_tensor).bool().to(tensor_stream.device).bool()
|
| 894 |
+
per_modality_stream = group_streams(tensor_stream.flat_stream(), group_fn=lambda ev: ev.type, schedule=False)
|
| 895 |
+
per_modality_compact_stream = {k: v.compact() for k, v in per_modality_stream.items()}
|
| 896 |
+
|
| 897 |
+
# Fill 'data' and 'set_data_mask' for each requested stream type
|
| 898 |
+
for st in stream_types:
|
| 899 |
+
data_mask = st_tensor == st.value
|
| 900 |
+
partial_mask = (
|
| 901 |
+
per_modality_stream[st]
|
| 902 |
+
.map_compact(
|
| 903 |
+
lambda ev: [int(ev.idx_range[0] <= i < ev.idx_range[1]) for i in range(ev.num_tokens(partial=False))]
|
| 904 |
+
)
|
| 905 |
+
.bool()
|
| 906 |
+
)
|
| 907 |
+
data[data_mask] = per_modality_compact_stream[st].reshape(-1)[partial_mask]
|
| 908 |
+
|
| 909 |
+
roles_mask = torch.zeros_like(st_tensor).bool().to(tensor_stream.device).bool()
|
| 910 |
+
for role in roles:
|
| 911 |
+
roles_mask |= roles_tensor == ROLE_TO_IDX[role]
|
| 912 |
+
data_mask = data_mask & roles_mask
|
| 913 |
+
set_data_mask[data_mask] = True
|
| 914 |
+
|
| 915 |
+
return data, set_data_mask
|
| 916 |
+
|
| 917 |
+
|
| 918 |
+
def ts_slice(tensor_stream: TensorStream, start: int, end: int) -> TensorStream:
|
| 919 |
+
"""
|
| 920 |
+
Return a new TensorStream that contains *only* the tokens in the
|
| 921 |
+
half-open interval ``[start, end)`` (0-based, inclusive-exclusive).
|
| 922 |
+
"""
|
| 923 |
+
B, T = tensor_stream.shape
|
| 924 |
+
assert 0 <= start <= end <= T, f"slice [{start}, {end}) is out of bounds for sequence length {T}"
|
| 925 |
+
|
| 926 |
+
sliced_streams: list[Stream] = []
|
| 927 |
+
|
| 928 |
+
for stream in tensor_stream.streams:
|
| 929 |
+
# current position in tensor stream token dims
|
| 930 |
+
curr_global_index = 0
|
| 931 |
+
new_events: list[Event] = []
|
| 932 |
+
|
| 933 |
+
# iterate over each of the events in the stream only selecting
|
| 934 |
+
# the events that fall within the range
|
| 935 |
+
for ev in stream:
|
| 936 |
+
ev_len = ev.num_tokens()
|
| 937 |
+
|
| 938 |
+
# ev_start, ev_end are the start and end indicies of the
|
| 939 |
+
# event within the tensor stream token dim
|
| 940 |
+
global_ev_start, global_ev_end = curr_global_index, curr_global_index + ev_len
|
| 941 |
+
|
| 942 |
+
if global_ev_end <= start:
|
| 943 |
+
# The event occurs before the start skip it and move the cursor
|
| 944 |
+
# forward
|
| 945 |
+
curr_global_index = global_ev_end
|
| 946 |
+
continue
|
| 947 |
+
if global_ev_start >= end:
|
| 948 |
+
# event occurs after the end we can exit
|
| 949 |
+
break
|
| 950 |
+
|
| 951 |
+
# only consider the part of the event that falls within the range
|
| 952 |
+
keep_from = max(0, start - global_ev_start)
|
| 953 |
+
keep_to = min(ev_len, end - global_ev_start)
|
| 954 |
+
part = ev.shallow_copy()
|
| 955 |
+
|
| 956 |
+
if keep_from == 0 and keep_to == ev_len:
|
| 957 |
+
# Event lies wholly inside the slice
|
| 958 |
+
new_events.append(part)
|
| 959 |
+
else:
|
| 960 |
+
# Partial overlap → trim.
|
| 961 |
+
assert ev.is_measured
|
| 962 |
+
|
| 963 |
+
# update the local event ranges for the slices
|
| 964 |
+
sliced_event_start = part.idx_range[0] + keep_from
|
| 965 |
+
sliced_event_end = part.idx_range[0] + keep_to
|
| 966 |
+
part.slice_tokens(sliced_event_start, sliced_event_end)
|
| 967 |
+
new_events.append(part)
|
| 968 |
+
|
| 969 |
+
curr_global_index = global_ev_end
|
| 970 |
+
|
| 971 |
+
sliced_streams.append(create_stream(new_events, stream.priority, schedule=False))
|
| 972 |
+
|
| 973 |
+
return TensorStream(sliced_streams)
|
| 974 |
|
| 975 |
|
| 976 |
class PixelShuffleSiglip2VisionConfig(Siglip2VisionConfig):
|
|
|
|
| 1396 |
# Configuration
|
| 1397 |
# ============================================================================
|
| 1398 |
|
| 1399 |
+
MAX_PIXELS = 60_000_000 # 60-megapixel ceiling ≈ 8200 × 7300 px
|
| 1400 |
|
| 1401 |
# Vision preprocessing constants
|
| 1402 |
VISION_MEAN = (0.5, 0.5, 0.5)
|
|
|
|
| 1413 |
if arr.flags.writeable:
|
| 1414 |
return arr
|
| 1415 |
|
| 1416 |
+
# First, try the cheap path — in-place flag toggle (works for mmap'd arrays
|
| 1417 |
# and some shared memory buffers):
|
| 1418 |
try:
|
| 1419 |
arr.setflags(write=True)
|
| 1420 |
return arr # success: no data copy
|
| 1421 |
except ValueError:
|
| 1422 |
+
# Buffer is inherently read-only (e.g. backed by PyAV / PIL): make copy
|
| 1423 |
return arr.copy()
|
| 1424 |
|
| 1425 |
|
|
|
|
| 2545 |
"IsaacModel",
|
| 2546 |
"IsaacForConditionalGeneration",
|
| 2547 |
"IsaacProcessor",
|
| 2548 |
+
]
|