9results about How to "Reasonable location" patented technology

The present disclosure provides a time division duplex based semantic data transmission method and related equipment. The method comprises: a base station modulating and processing first user data corresponding to semantic data to obtain a first modulation sequence; the base station determining a first number of time-frequency resource blocks required for mapping the first modulation sequence to a first time-frequency resource; the base station judging whether the mapping of the first modulation sequence to the first time-frequency resource spans a time slot to obtain a first time slot span judgment result; the base station mapping the first modulation sequence to a first time-frequency resource with a first number of time-frequency resource blocks based on the first time slot span judgment result to obtain a first time-frequency signal, and sending the first time-frequency signal to a user terminal through a downlink; and the user terminal mapping the second modulation sequence to a second time-frequency resource with a second number of time-frequency resource blocks based on a second time slot span judgment result to obtain a second time-frequency signal, and sending the second time-frequency signal to the base station through an uplink.

The application discloses a paper feeding cutter device which comprises a frame body, a driving device, a pressing assembly, a transmission assembly and a paper separating wheel installed on the frame body, the pressing assembly is used for pressing paper in an elastic manner, the paper separating wheel is arranged adjacent to the pressing assembly, the paper separating wheel is rotatably installed on the frame body, and the paper separating wheel is further installed on an output end of the driving device through the transmission assembly, so that the paper separating wheel rotates under the driving of the driving device, and the paper separating wheel can separate overlapped paper and make the paper pass in a single sheet mode through the rotation of the paper separating wheel. The paper feeding cutter device can effectively separate paper and avoid thin paper stacking, and has the advantages of simple structure.

The utility model discloses a pearl sucking mechanism belongs to pearl processing equipment technical field, and this mechanism mainly includes puncher main part, and below has drive motor and material taking cavity, and the inside operating table top has straight vibration platform, feed hopper, pearl clamp, CCD camera and punch device etc., and the operating table top still is equipped with array stand, and the pearl sucking mechanism main part is pearl sucking mechanism main part in the stand inboard, and the stand crossbeam top has longitudinal slide rail and horizontal slide rail, and the outside has longitudinal drive motor, and through the drag chain connection ensures the smooth operation, and the pearl sucking mechanism main part front side has the mounting panel, the sucking mechanism motor, the fixed disc etc., and the fixed disc front side has the steering motor, and the bottom passes through the pivot, the buffer spring and connects the sucking disc, and the steering motor can adjust pearl position according to CCD camera, and the puncher main part bottom has the moving wheel and the positioning disc, and the convenient removal and fixed, this pearl sucking mechanism convenient operation, high precision, good stability can accurate positioning and sucking pearl, reduce the damage to pearl, improve the processing efficiency and precision.

The utility model discloses a modularization ship living cabin relates to ocean engineering technical field, the utility model discloses a cabin, the cabin inner wall rotation is connected with the cabin door, be provided with ventilation mechanism on the cabin, be provided with dismounting mechanism on the cabin, the ventilation mechanism includes the air intake, the cabin outer surface sliding connection has the connecting cover, the connecting cover inner wall screw thread connection has the bolt no.

This disclosure provides a semantic data transmission method and related equipment based on frequency division duplex (FDM). The method includes: a base station modulating first user data corresponding to semantic data to obtain a first modulation sequence; the base station determining a first target combination method of the first modulation sequence based on the semantic type of the semantic data; the base station performing time-frequency resource mapping processing on the first modulation sequence based on the first target combination method to obtain a first time-frequency signal, and transmitting the first time-frequency signal to a user terminal via a downlink; the user terminal modulating second user data corresponding to the semantic data to obtain a second modulation sequence; the user terminal determining a second target combination method of the second modulation sequence based on the semantic type of the semantic domain information; the user terminal performing time-frequency resource mapping processing on the second modulation sequence based on the second target combination method to obtain a second time-frequency signal, and transmitting the second time-frequency signal to the base station via an uplink.

The application discloses a specimen classification sending device, which comprises a specimen temporary storage bin, a standardization mechanism, a code scanning mechanism, a pushing mechanism, a direction adjusting mechanism, a specimen sending module and a multi-pipeline path switching device which are sequentially connected. The pushing mechanism comprises a guide slide rail, a specimen sliding groove arranged on one side of the guide slide rail, a pushing top rod slidingly arranged on the guide slide rail and a first driving mechanism for driving the pushing top rod to reciprocate linearly. An end of the pushing top rod is provided with a top rod contact, the top rod contact is arranged in the specimen sliding groove, an arc-shaped limiting groove is formed in the bottom of the specimen sliding groove, the diameter of the arc-shaped limiting groove is larger than the diameter of a small head of the specimen and smaller than the diameter of a large head of the specimen, a conical opening is formed in the top rod contact, a reflective infrared sensor and a magnetic steel arranged on the side of the top rod contact are further arranged on the top rod contact, and the center of the reflective infrared sensor is coincident with the center of the opening.

This invention discloses a joint deployment optimization method for hybrid expert model inference in heterogeneous GPU environments. The method first uses offline trajectory statistics to determine expert activation frequencies and collects resource profiles of heterogeneous devices. Then, it constructs a joint optimization model with the objective of minimizing the weighted sum of inference latency surrogate terms and quantization error penalty terms. Under memory constraints, a pruned greedy search algorithm is used to solve for the placement location, number, and quantization accuracy of each expert's replicas. During the online inference phase, a routing mechanism combining local priority and load-aware weighted round-robin is used for dynamic token distribution. This invention can simultaneously address inference latency, memory usage, and accuracy preservation in heterogeneous GPU environments, improving the processing efficiency of popular experts, reducing cross-device communication overhead and slow device trailing effects, and is suitable for distributed large-scale model inference deployment scenarios.

This invention discloses a joint deployment optimization method for hybrid expert model inference in heterogeneous GPU environments. The method first uses offline trajectory statistics to determine expert activation frequencies and collects resource profiles of heterogeneous devices. Then, it constructs a joint optimization model with the objective of minimizing the weighted sum of inference latency surrogate terms and quantization error penalty terms. Under memory constraints, a pruned greedy search algorithm is used to solve for the placement location, number, and quantization accuracy of each expert's replicas. During the online inference phase, a routing mechanism combining local priority and load-aware weighted round-robin is used for dynamic token distribution. This invention can simultaneously address inference latency, memory usage, and accuracy preservation in heterogeneous GPU environments, improving the processing efficiency of popular experts, reducing cross-device communication overhead and slow device trailing effects, and is suitable for distributed large-scale model inference deployment scenarios.

The utility model relates to an entertainment facility, disclose a kind of supporting leg subassembly, pedestal and card machine. Card machine includes pedestal, and supporting leg subassembly is equipped on the pedestal, and supporting leg subassembly includes supporting leg and the buffer assembly for realizing indirect contact between supporting leg and pedestal, which is installed on the pedestal, the buffer assembly is used to provide resistance when supporting leg and the relative motion of the main body of pedestal at the cooperation stroke;Buffer assembly includes two buffer springs, which are arranged on the pedestal and are compressed on both sides of the supporting leg respectively. The utility model sets up buffer assembly on supporting leg subassembly, which can not only play a better buffering and damping effect in the process of rotation start and stop, but also avoid interference to the internal transmission components of card machine, with reasonable position, stable work and other characteristics.