Giorgio Oppo
0231312820
Share the established M1 Max dispatch policy with the M2-M4 families without changing kernel arithmetic, supported tensor shapes, math-mode guards or diagnostic overrides. - Centralize device eligibility with generation-bounded name matching. Preserve M1 base/Pro/Ultra and M5/M6 defaults, and retain M3 Ultra's dedicated MoE specialization, geometry and prefill tile policy. - Enable wide Q8 decode row reuse, HC gate/mix reuse, short MTP Q8 projections and the existing IQ2/Q2/MXFP4 expert paths on M2-M4. - Extend small HC prefill tiles, large Q8 SSD prefill, sparse IQ2/Q2 SSD tiles, Q4/MXFP4 K32 tiles, large-batch half staging and Q4 decode scratch release under the same existing shape and batch guards. - Update SSD dispatch probes, cover 47 device-name cases and document the supported families and the limits of current hardware evidence. Validation: build all application targets and run the device-policy, Qwen kernel, safe-math reuse/projection, MoE specialization and SSD expert tests (retained and unretained command buffers) on an isolated HEAD export containing only this commit. Local hardware is M1 Max; M2/M3/M4 throughput still requires validation on those devices.
32 lines
1.3 KiB
C
32 lines
1.3 KiB
C
#ifndef DS4_METAL_DEVICE_H
|
|
#define DS4_METAL_DEVICE_H
|
|
|
|
#include <string.h>
|
|
|
|
/* Share the Qwen M1 Max defaults with M2-M4 while keeping other M1 devices
|
|
* and newer generations on their existing paths. Bound the generation token
|
|
* so names such as M20 do not inherit these defaults. */
|
|
static inline int ds4_metal_device_name_uses_qwen_legacy_tuning(const char *name) {
|
|
return name && (strcmp(name, "Apple M1 Max") == 0 ||
|
|
(strncmp(name, "Apple M", 7) == 0 &&
|
|
name[7] >= '2' && name[7] <= '4' &&
|
|
(name[8] == '\0' || name[8] == ' ')));
|
|
}
|
|
|
|
/* M3 Ultra already has dedicated MoE defaults; preserve those selections. */
|
|
static inline int ds4_metal_device_name_uses_qwen_legacy_moe_tuning(const char *name) {
|
|
return ds4_metal_device_name_uses_qwen_legacy_tuning(name) &&
|
|
strcmp(name, "Apple M3 Ultra") != 0;
|
|
}
|
|
|
|
/* Share the M5 tuning policy with M6. Keep the generation token bounded so
|
|
* untested future names such as M50 and M60 do not inherit these defaults.
|
|
* Callers must still check runtime and kernel capabilities where required. */
|
|
static inline int ds4_metal_device_name_is_m5_or_m6(const char *name) {
|
|
return name && strncmp(name, "Apple M", 7) == 0 &&
|
|
(name[7] == '5' || name[7] == '6') &&
|
|
(name[8] == '\0' || name[8] == ' ');
|
|
}
|
|
|
|
#endif
|