Fix CUDA vision build

This commit is contained in:
antirez committed 2026-09-01 00:32:06 +02:00
1 parent 409080a623
commit 200b9f45bf
1 file changed
+36
+36
View File
@@ -1,5 +1,41 @@
/* DeepSeek V4 Flash Vision-Exp operations shared by CUDA and ROCm. */
#ifndef DS4_DEEPSEEK4_VISION_TYPES_DEFINED
#define DS4_DEEPSEEK4_VISION_TYPES_DEFINED
#define DS4_DEEPSEEK4_VISION_LAYERS 32u
#define DS4_DEEPSEEK4_LANGUAGE_LAYERS 43u
#define DS4_DEEPSEEK4_MTP_LAYERS 3u
typedef struct {
uint64_t norm1;
uint64_t qkv_weight;
uint64_t qkv_bias;
uint64_t attn_proj_weight;
uint64_t attn_proj_bias;
uint64_t norm2;
uint64_t mlp_w1;
uint64_t mlp_w2;
} ds4_deepseek4_vision_layer_weights;
typedef struct {
uint64_t patch_weight;
uint64_t patch_bias;
uint64_t post_norm;
uint64_t aligner_w1;
uint64_t aligner_w1_bias;
uint64_t aligner_w2;
uint64_t aligner_w2_bias;
uint64_t image_start;
uint64_t image_pad;
uint64_t image_newline;
uint64_t image_end;
uint64_t visual_router_bias[DS4_DEEPSEEK4_LANGUAGE_LAYERS];
uint64_t mtp_visual_router_bias[DS4_DEEPSEEK4_MTP_LAYERS];
uint64_t hash_router_bias[3];
ds4_deepseek4_vision_layer_weights layer[DS4_DEEPSEEK4_VISION_LAYERS];
} ds4_deepseek4_vision_weights;
#endif
#ifndef DS4_DEEPSEEK4_VISION_STREAM
#define DS4_DEEPSEEK4_VISION_STREAM 0
#endif