Abstract: Methods, systems, and apparatus for propagating reduced-precision on computation graphs are described. In one aspect, a method includes receiving data specifying a directed graph that includes operators for a program. The operators include first operators that each represent a numerical operation performed on numerical values having a first level of precision and second operators that each represent a numerical operation performed on numerical values having a second level of precision. One or more downstream operators are identified for a first operator. A determination is made whether each downstream operator represents a numerical operation that is performed on input values having the second level of precision. Whenever each downstream operator represents a numerical operation that is performed on input values having the second level of precision, a precision of numerical values output by the operation represented by the first operator is adjusted to the second level of precision.

Abstract: Systems and methods are provided for vector-quantized image modeling using vision transformers and improved codebook handling. In particular, the present disclosure provides a Vector-quantized Image Modeling (VIM) approach that involves pretraining a machine learning model (e.g., Transformer model) to predict rasterized image tokens autoregressively. The discrete image tokens can be encoded from a learned Vision-Transformer-based VQGAN (example implementations of which can be referred to as ViT-VQGAN). The present disclosure proposes multiple improvements over vanilla VQGAN from architecture to codebook learning, yielding better efficiency and reconstruction fidelity. The improved ViT-VQGAN further improves vector-quantized image modeling tasks, including unconditional image generation, conditioned image generation (e.g., class-conditioned image generation), and unsupervised representation learning.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for performing machine learning task on a network input to generate a network output. In one aspect, one of the systems includes an attention neural network configured to perform the machine learning task, the attention neural network including one or more attention layers, each attention layer comprising an attention sub-layer and a feed-forward sub-layer. Some or all of the attention layers have a feed-forward sub-layer that applies conditional computation to the inputs to the sub-layer.

Abstract: Methods, systems, and apparatus for propagating reduced-precision on computation graphs are described. In one aspect, a method includes receiving data specifying a directed graph that includes operators for a program. The operators include first operators that each represent a numerical operation performed on numerical values having a first level of precision and second operators that each represent a numerical operation performed on numerical values having a second level of precision. One or more downstream operators are identified for a first operator. A determination is made whether each downstream operator represents a numerical operation that is performed on input values having the second level of precision. Whenever each downstream operator represents a numerical operation that is performed on input values having the second level of precision, a precision of numerical values output by the operation represented by the first operator is adjusted to the second level of precision.

Abstract: Methods, systems, and apparatus for propagating reduced-precision on computation graphs are described. In one aspect, a method includes receiving data specifying a directed graph that includes operators for a program. The operators include first operators that each represent a numerical operation performed on numerical values having a first level of precision and second operators that each represent a numerical operation performed on numerical values having a second level of precision. One or more downstream operators are identified for a first operator. A determination is made whether each downstream operator represents a numerical operation that is performed on input values having the second level of precision. Whenever each downstream operator represents a numerical operation that is performed on input values having the second level of precision, a precision of numerical values output by the operation represented by the first operator is adjusted to the second level of precision.

Abstract: Methods, systems, and apparatus for scheduling first-in-first-out instructions are described. In one aspect, a method includes receiving data representing code of a program to be executed by a processing unit comprising hardware processors. For each of one or more of the hardware processors, an order of independent groups of first-in-first-out (FIFO) instructions for execution by the hardware processor is identified in the data representing the code of the program. For each independent group of FIFO instructions for execution by the hardware processor, a path length metric that represents how long it will take to reach an end of the program from the independent group of FIFO instructions is determined. A new order of the independent groups of FIFO instructions for execution by the hardware processor is generated based at least on the path length metric for each independent group of FIFO instructions for execution by the hardware processor.

Abstract: Methods, systems, and apparatus for scheduling first-in-first-out instructions are described. In one aspect, a method includes receiving data representing code of a program to be executed by a processing unit comprising hardware processors. For each of one or more of the hardware processors, an order of independent groups of first-in-first-out (FIFO) instructions for execution by the hardware processor is identified in the data representing the code of the program. For each independent group of FIFO instructions for execution by the hardware processor, a path length metric that represents how long it will take to reach an end of the program from the independent group of FIFO instructions is determined. A new order of the independent groups of FIFO instructions for execution by the hardware processor is generated based at least on the path length metric for each independent group of FIFO instructions for execution by the hardware processor.

Abstract: Methods, systems, and apparatus for propagating reduced-precision on computation graphs are described. In one aspect, a method includes receiving data specifying a directed graph that includes operators for a program. The operators include first operators that each represent a numerical operation performed on numerical values having a first level of precision and second operators that each represent a numerical operation performed on numerical values having a second level of precision. One or more downstream operators are identified for a first operator. A determination is made whether each downstream operator represents a numerical operation that is performed on input values having the second level of precision. Whenever each downstream operator represents a numerical operation that is performed on input values having the second level of precision, a precision of numerical values output by the operation represented by the first operator is adjusted to the second level of precision.

Abstract: Methods, systems, and apparatus for scheduling first-in-first-out instructions are described. In one aspect, a method includes receiving data representing code of a program to be executed by a processing unit comprising hardware processors. For each of one or more of the hardware processors, an order of independent groups of first-in-first-out (FIFO) instructions for execution by the hardware processor is identified in the data representing the code of the program. For each independent group of FIFO instructions for execution by the hardware processor, a path length metric that represents how long it will take to reach an end of the program from the independent group of FIFO instructions is determined. A new order of the independent groups of FIFO instructions for execution by the hardware processor is generated based at least on the path length metric for each independent group of FIFO instructions for execution by the hardware processor.

Abstract: An integrated lens with integrated functional optical surfaces for use in optical communication is disclosed. The integrated lens includes first and second cavities and a fiber adapter. The device also includes integrated first and second lenses. The first cavity houses one or more optical transmitting and/or receiving devices. The second cavity has a first optical surface and an optional second optical surface. The fiber adapter has the second lens. The integrated lens enables a small size, a light weight, high coupling and a high transmission efficiency, and can be produced by injection molding using a single mold. The integrated lens is applicable to optical signal coupling, fiber connections, optical modules, and optical or optoelectronic communication.

Abstract: An integrated lens with integrated functional optical surfaces for use in optical communication is disclosed. The integrated lens includes first and second cavities and a fiber adapter. The device also includes integrated first and second lenses. The first cavity houses one or more optical transmitting and/or receiving devices. The second cavity has a first optical surface and an optional second optical surface. The fiber adapter has the second lens. The integrated lens enables a small size, a light weight, high coupling and a high transmission efficiency, and can be produced by injection molding using a single mold. The integrated lens is applicable to optical signal coupling, fiber connections, optical modules, and optical or optoelectronic communication.

Abstract: An N×N parallel optical transceiver includes a printed circuit board, a laser driving control chip, one or more lasers, two GRIN lenses, an optical band-pass filter, a multimode fiber array and a photodiode array. In the transmitter, laser beams of the same wavelength simultaneously output from the laser chip are first focused by the first GRIN lens, then the beams pass through a wavelength band-pass filter and are refocused by the second GRIN lens, and enter the channels in the multimode fiber array. In the receiver, laser beams of a different wavelength from the multimode fiber array are focused by the second GRIN lens, then reflected by the band-pass filter, refocused by the second GRIN lens, and received by the photodiode array. The multi-channel parallel transceiver has a small form, and can integrate a DFB or FP laser chip and GRIN lenses.

Abstract: A dual-rate power point compensating circuit, comprising a microprocessor and a transmitter optical subsystem assembly (TOSA), wherein the TOSA includes a laser connected to a laser driver, a monitor photodiode (MPD) connected to the laser driver, and a current divider connected to the microprocessor and the MPD. When a feedback current from the MPD exceeds the adjustable and/or operating range of the laser driver, the feedback current is reduced so that it is kept in the adjustable and/or operating range of the laser driver. The laser driver determines the optical output power of the laser from the value of the reduced feedback current. The circuit and method extend the adjustable and/or operating range of a laser driver and enable it to regulate a target optical output power of the laser with a broad testing range and high accuracy when the feedback current is relatively high.

Abstract: A DC level detection circuit between high speed signal line connecting ports and a system using the same in optical communications is disclosed. Corresponding ports of a high speed signal line each have an additional resistor, where one additional resistor has a resistance significantly greater than that of the other resistor. The smaller resistor is grounded. The larger resistor is connected to a DC voltage source, a low pass filter, and a signal detection port. Thus, when both ports of the high speed signal line are connected, a status of the electrical level detected at the signal detection port changes. The circuit detects the connection state of the high speed signal line without negative effects on signal transmission and is applicable to various circuits, especially plug-in modules and corresponding slots. Thus, the circuit further enables signal port multifunctionality, increases module installation accuracy, and provides higher compatibility for plug-in modules.

Abstract: The present application discloses a method and circuitry that improves the monitoring and/or reporting accuracy and of a TOSA transmitter output power. In the method, the output power of an optical transmitter is measured at 25° C. and at N individual temperatures to obtain N tracking error (TE) values corresponding to the N individual temperature values, then a lookup table covering an operating temperature range of the transmitter is created based on a one-to-one mapping relationship between the TE values and the N individual temperatures and a line fitting process. The transmitter output power is reported at an interface of the transmitter according to the TE value at the transmitter operating temperature in the lookup table. The present application also discloses optical modules and optical communication systems. The present method and transmitter effectively improve the monitoring and/or reporting accuracy of the transmitter output power.

Abstract: A gain-variable trans-impedance amplifier (TIA) in optical device is disclosed. The TIA has an improved dynamic range for receiving electrical signals and is configured to convert current signals from an avalanche photodiode (APD) to voltage signals. A resistor element is between the input and output terminals of the TIA, wherein the resistance of the resistor element can regulate the resistance and/or impedance value of the TIA, and a switch determines or controls the resistance of the resistor element. When the power of an optical signal received by the APD is higher than a predetermined value, the resistance becomes smaller and the gain of the TIA becomes greater. When the power of the optical signal is lower than the predetermined value, the resistance becomes greater. The gain of the TIA is automatically adjusted on the basis of the intensity of received optical signals to obtain a greater dynamic operational range.

Abstract: A high speed optical module, and in particular, a multi-channel, single-mode, parallel transmission optical module in the field of optical communication is disclosed. The optical module includes a chassis, a first transmitter optical subassembly (TOSA), a second transmitter optical subassembly (TOSA), a third transmitter optical subassembly (TOSA), a fourth transmitter optical subassembly (TOSA) and a MT fiber connector. The TOSA may be a TO-38 TOSA of a 10 G DFB chip or FP chip. With single-mode communication, the optical module provides a transmission distance over 10 kilometers. In addition, to make coupling single-mode fibers easier, a twelve-core MT fiber connector is employed, wherein four fiber connectors are respectively connected to LC standard fiber stubs and the outputs of the TOSAs accordingly. Thus, the optical module further provides high speed and long-distance transmission, large capacity, small volume, and light weight and can be broadly applied to high speed optical communications.

Abstract: An N×N parallel optical transceiver includes a printed circuit board, a laser driving control chip, one or more lasers, two GRIN lenses, an optical band-pass filter, a multimode fiber array and a photodiode array. In the transmitter, laser beams of the same wavelength simultaneously output from the laser chip are first focused by the first GRIN lens, then the beams pass through a wavelength band-pass filter and are refocused by the second GRIN lens, and enter the channels in the multimode fiber array. In the receiver, laser beams of a different wavelength from the multimode fiber array are focused by the second GRIN lens, then reflected by the band-pass filter, refocused by the second GRIN lens, and received by the photodiode array. The multi-channel parallel transceiver has a small form, and can integrate a DFB or FP laser chip and GRIN lenses.

Abstract: An integrated lens with integrated functional optical surfaces for use in optical communication is disclosed. The integrated lens includes first and second cavities and a fiber adapter. The device also includes integrated first and second lenses. The first cavity houses one or more optical transmitting and/or receiving devices. The second cavity has a first optical surface and an optional second optical surface. The fiber adapter has the second lens. The integrated lens enables a small size, a light weight, high coupling and a high transmission efficiency, and can be produced by injection molding using a single mold. The integrated lens is applicable to optical signal coupling, fiber connections, optical modules, and optical or optoelectronic communication.

Abstract: A gain-variable trans-impedance amplifier (TIA) in optical device is disclosed. The TIA has an improved dynamic range for receiving electrical signals and is configured to convert current signals from an avalanche photodiode (APD) to voltage signals. A resistor element is between the input and output terminals of the TIA, wherein the resistance of the resistor element can regulate the resistance and/or impedance value of the TIA, and a switch determines or controls the resistance of the resistor element. When the power of an optical signal received by the APD is higher than a predetermined value, the resistance becomes smaller and the gain of the TIA becomes greater. When the power of the optical signal is lower than the predetermined value, the resistance becomes greater. The gain of the TIA is automatically adjusted on the basis of the intensity of received optical signals to obtain a greater dynamic operational range.