antirez
43a91aaa3b
Split the drafter and verifier vocabulary head across peers, negotiate their state and memory requirements, and retain local drafting when the split is unavailable. Use an ordered RDMA command mailbox for verification and commits. Preserve accepted-prefix history through restores and skips. Balance MXFP4 experts, stage short expert blocks, overlap independent projections and shared experts, batch compressor and frontier copies, and specialize measured Q8 projection shapes. Fix active-stream ordering, paired projection row coverage and the model's indexer top-k threshold. Add independent numerical references, graph and stream tests, transport failure tests, and speculative session lifecycle coverage. The fused HC path preserves the mathematical graph but changes rounding; official continuation scores are mixed, not universally improved. A pre-existing scalar-versus-batched book oracle discrepancy remains under investigation.
25 lines
639 B
C
25 lines
639 B
C
#ifndef DS4_GPU_COPY_H
|
|
#define DS4_GPU_COPY_H
|
|
|
|
#include <stdint.h>
|
|
|
|
/* Device copy spans, shared by ds4_gpu.h and the CUDA backend, which does
|
|
* not include ds4_gpu.h. The tensor stays opaque here. */
|
|
#ifndef DS4_GPU_TENSOR_DEFINED
|
|
#define DS4_GPU_TENSOR_DEFINED
|
|
typedef struct ds4_gpu_tensor ds4_gpu_tensor;
|
|
#endif
|
|
|
|
/* One device copy, with the arguments of ds4_gpu_tensor_copy. */
|
|
typedef struct {
|
|
ds4_gpu_tensor *dst;
|
|
uint64_t dst_offset;
|
|
const ds4_gpu_tensor *src;
|
|
uint64_t src_offset;
|
|
uint64_t bytes;
|
|
} ds4_gpu_copy_span;
|
|
|
|
#define DS4_GPU_COPY_SPANS_MAX 316u
|
|
|
|
#endif
|