server : fix mtmd checkpoints (#16591)

* metal : avoid using Metal's gpuAddress property

* metal : fix rope kernels buffer check

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -2876,7 +2876,7 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
         }
 
         //if rms norm is the B operand, then we don't handle broadcast
-        if (rms_norm == mul->src[1] && !ggml_are_same_shape(mul->src[0], rms_norm->src[1])) {
+        if (rms_norm == mul->src[1] && !ggml_are_same_shape(mul->src[0], rms_norm)) {
             return false;
         }
 

ggml/src/ggml-metal/ggml-metal-device.m

@@ -7,6 +7,8 @@
 
 #include <Metal/Metal.h>
 
+#include <stdatomic.h>
+
 #ifndef TARGET_OS_VISION
 #define TARGET_OS_VISION 0
 #endif
@@ -22,6 +24,9 @@
 // overload of MTLGPUFamilyMetal3 (not available in some environments)
 static const NSInteger MTLGPUFamilyMetal3_GGML = 5001;
 
+// virtual address for GPU memory allocations
+static atomic_uintptr_t g_addr_device = 0x000000400ULL;
+
 #if !GGML_METAL_EMBED_LIBRARY
 // Here to assist with NSBundle Path Hack
 @interface GGMLMetalClass : NSObject
@@ -827,7 +832,7 @@ struct ggml_metal_buffer_wrapper {
 };
 
 struct ggml_metal_buffer {
-    void * all_data; // TODO: https://github.com/ggml-org/llama.cpp/pull/15985
+    void * all_data;
     size_t all_size;
 
     // if false, the Metal buffer data is allocated in private GPU memory and is not shared with the host
@@ -965,14 +970,15 @@ ggml_metal_buffer_t ggml_metal_buffer_init(ggml_metal_device_t dev, size_t size,
     if (shared) {
         res->all_data = ggml_metal_host_malloc(size_aligned);
         res->is_shared = true;
-        res->owned = true;
     } else {
-        // dummy, non-NULL value - we'll populate this after creating the Metal buffer below
-        res->all_data = (void *) 0x000000400ULL;
+        // use virtual address from g_addr_device counter
+        res->all_data = (void *) atomic_fetch_add_explicit(&g_addr_device, size_aligned, memory_order_relaxed);
         res->is_shared = false;
     }
     res->all_size = size_aligned;
 
+    res->owned = true;
+
     res->device = ggml_metal_device_get_obj(dev);
     res->queue  = ggml_metal_device_get_queue(dev);
 
@@ -983,15 +989,13 @@ ggml_metal_buffer_t ggml_metal_buffer_init(ggml_metal_device_t dev, size_t size,
         res->buffers[0].metal = nil;
 
         if (size_aligned > 0) {
-            if (props_dev->use_shared_buffers &&shared) {
+            if (props_dev->use_shared_buffers && shared) {
                 res->buffers[0].metal = [res->device newBufferWithBytesNoCopy:res->all_data
                                                                   length:size_aligned
                                                                  options:MTLResourceStorageModeShared
                                                              deallocator:nil];
             } else {
                 res->buffers[0].metal = [res->device newBufferWithLength:size_aligned options:MTLResourceStorageModePrivate];
-
-                res->all_data = (void *) (res->buffers[0].metal.gpuAddress);
             }
         }
 
@@ -1139,7 +1143,7 @@ bool ggml_metal_buffer_is_shared(ggml_metal_buffer_t buf) {
 
 void ggml_metal_buffer_memset_tensor(ggml_metal_buffer_t buf, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
     if (buf->is_shared) {
-        memset((char *)tensor->data + offset, value, size);
+        memset((char *) tensor->data + offset, value, size);
         return;
     }
 
@@ -1168,7 +1172,7 @@ void ggml_metal_buffer_memset_tensor(ggml_metal_buffer_t buf, struct ggml_tensor
 
 void ggml_metal_buffer_set_tensor(ggml_metal_buffer_t buf, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
     if (buf->is_shared) {
-        memcpy((char *)tensor->data + offset, data, size);
+        memcpy((char *) tensor->data + offset, data, size);
         return;
     }
 
@@ -1223,7 +1227,7 @@ void ggml_metal_buffer_set_tensor(ggml_metal_buffer_t buf, struct ggml_tensor *
 
 void ggml_metal_buffer_get_tensor(ggml_metal_buffer_t buf, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
     if (buf->is_shared) {
-        memcpy(data, (const char *)tensor->data + offset, size);
+        memcpy(data, (const char *) tensor->data + offset, size);
         return;
     }
 

ggml/src/ggml-metal/ggml-metal-impl.h

@@ -251,6 +251,7 @@ typedef struct {
     int32_t  sect_1;
     int32_t  sect_2;
     int32_t  sect_3;
+    bool     src2;
 } ggml_metal_kargs_rope;
 
 typedef struct {

ggml/src/ggml-metal/ggml-metal-ops.cpp

@@ -2969,6 +2969,7 @@ int ggml_metal_op_rope(ggml_metal_op_t ctx, int idx) {
         /* sect_1      =*/ sect_1,
         /* sect_2      =*/ sect_2,
         /* sect_3      =*/ sect_3,
+        /* src2        =*/ op->src[2] != nullptr,
     };
 
     ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_rope(lib, op);

ggml/src/ggml-metal/ggml-metal.metal

@@ -3748,7 +3748,7 @@ kernel void kernel_rope_norm(
 
             const float theta = theta_base * pow(args.freq_base, inv_ndims*i0);
 
-            const float freq_factor = src2 != src0 ? ((device const float *) src2)[ic] : 1.0f;
+            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;
 
             rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
 
@@ -3801,7 +3801,7 @@ kernel void kernel_rope_neox(
 
             const float theta = theta_base * pow(args.freq_base, inv_ndims*i0);
 
-            const float freq_factor = src2 != src0 ? ((device const float *) src2)[ic] : 1.0f;
+            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;
 
             rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
 
@@ -3872,7 +3872,7 @@ kernel void kernel_rope_multi(
 
             const float theta = theta_base * pow(args.freq_base, inv_ndims*i0);
 
-            const float freq_factor = src2 != src0 ? ((device const float *) src2)[ic] : 1.0f;
+            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;
 
             rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
 
@@ -3939,7 +3939,7 @@ kernel void kernel_rope_vision(
             const float theta = theta_base * pow(args.freq_base, 2.0f * inv_ndims * p);
             // end of mrope
 
-            const float freq_factor = src2 != src0 ? ((device const float *) src2)[ic] : 1.0f;
+            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;
 
             rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
 

ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -2686,7 +2686,7 @@ static bool ggml_opencl_can_fuse(const struct ggml_cgraph * cgraph, int node_idx
 
         // if rms_norm is the B operand, then we don't handle broadcast
         if (rms_norm == mul->src[1] &&
-            !ggml_are_same_shape(mul->src[0], rms_norm->src[1])) {
+            !ggml_are_same_shape(mul->src[0], rms_norm)) {
             return false;
         }
 

ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp

@@ -313,12 +313,12 @@ void main() {
         sums[i] = coopmat<ACC_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator>(0.0f);
     }
 #else
-    ACC_TYPE sums[WMITER * TM * WNITER * TN];
+    ACC_TYPE_VEC2 sums[WMITER * TM * WNITER * TN/2];
     FLOAT_TYPE_VEC2 cache_a[WMITER * TM];
-    FLOAT_TYPE_VEC2 cache_b[TN];
+    FLOAT_TYPE_VEC2 cache_b;
 
-    [[unroll]] for (uint i = 0; i < WMITER*TM*WNITER*TN; i++) {
-        sums[i] = ACC_TYPE(0.0f);
+    [[unroll]] for (uint i = 0; i < WMITER*TM*WNITER*TN/2; i++) {
+        sums[i] = ACC_TYPE_VEC2(0.0f, 0.0f);
     }
 #endif
 
@@ -360,20 +360,22 @@ void main() {
                     cache_a[wsir * TM + j] = buf_a[(warp_r * WM + wsir * WSUBM + tiwr * TM + j) * SHMEM_STRIDE + i];
                 }
             }
-            [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {
-                [[unroll]] for (uint j = 0; j < TN; j++) {
-                    cache_b[j] = buf_b[(warp_c * WN + wsic * WSUBN + tiwc * TN + j) * SHMEM_STRIDE + i];
-                }
 
-                [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
-                    [[unroll]] for (uint cc = 0; cc < TN; cc++) {
-                        [[unroll]] for (uint cr = 0; cr < TM; cr++) {
-                            const uint sums_idx = (wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr;
-                            sums[sums_idx] = fma(ACC_TYPE(cache_a[wsir * TM + cr].x), ACC_TYPE(cache_b[cc].x), fma(ACC_TYPE(cache_a[wsir * TM + cr].y), ACC_TYPE(cache_b[cc].y), sums[sums_idx]));
+            [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {
+                [[unroll]] for (uint cc = 0; cc < TN; cc++) {
+                    cache_b = buf_b[(warp_c * WN + wsic * WSUBN + tiwc * TN + cc) * SHMEM_STRIDE + i];
+
+                    [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
+                        [[unroll]] for (uint cr = 0; cr < TM / 2; cr++) {
+                            // [WNITER][TN][WMITER][TM / 2] -> [wsic][cc][wsir][cr]
+                            const uint sums_idx = (wsic * TN + cc) * WMITER * (TM / 2) + wsir * (TM / 2) + cr;
+                            sums[sums_idx].x = fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr    ].x), ACC_TYPE(cache_b.x), fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr    ].y), ACC_TYPE(cache_b.y), sums[sums_idx].x));
+                            sums[sums_idx].y = fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr + 1].x), ACC_TYPE(cache_b.x), fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr + 1].y), ACC_TYPE(cache_b.y), sums[sums_idx].y));
                         }
                     }
                 }
             }
+
         }
 #endif
 
@@ -388,8 +390,9 @@ void main() {
         }
     }
 #else
-    [[unroll]] for (uint i = 0; i < WMITER*TM*WNITER*TN; i++) {
-        sums[i] = clamp(sums[i], -ACC_TYPE_MAX, ACC_TYPE_MAX);
+    [[unroll]] for (uint i = 0; i < WMITER*TM*WNITER*TN/2; i++) {
+        sums[i].x = clamp(sums[i].x, -ACC_TYPE_MAX, ACC_TYPE_MAX);
+        sums[i].y = clamp(sums[i].y, -ACC_TYPE_MAX, ACC_TYPE_MAX);
     }
 #endif
 #endif
@@ -463,14 +466,21 @@ void main() {
 
                 const u16vec2 row_idx = row_ids[row_i - ic * BN];
 #endif // MUL_MAT_ID
-                [[unroll]] for (uint cr = 0; cr < TM; cr++) {
+                [[unroll]] for (uint cr = 0; cr < TM / 2; cr++) {
+                    const uint sums_idx = (wsic * TN + cc) * WMITER * (TM / 2) + wsir * (TM / 2) + cr;
 #ifdef MUL_MAT_ID
-                    if (dr_warp + cr < p.M) {
-                        data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]);
+                    if (dr_warp + 2 * cr < p.M) {
+                        data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr_warp + 2 * cr] = D_TYPE(sums[sums_idx].x);
+                    }
+                    if (dr_warp + 2 * cr + 1 < p.M) {
+                        data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr_warp + 2 * cr + 1] = D_TYPE(sums[sums_idx].y);
                     }
 #else
-                    if (dr_warp + cr < p.M && dc_warp + cc < p.N) {
-                        data_d[offsets + (dc_warp + cc) * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]);
+                    if (dr_warp + 2 * cr < p.M && dc_warp + cc < p.N) {
+                        data_d[offsets + (dc_warp + cc) * p.stride_d + dr_warp + 2 * cr] = D_TYPE(sums[sums_idx].x);
+                    }
+                    if (dr_warp + 2 * cr + 1 < p.M && dc_warp + cc < p.N) {
+                        data_d[offsets + (dc_warp + cc) * p.stride_d + dr_warp + 2 * cr + 1] = D_TYPE(sums[sums_idx].y);
                     }
 #endif // MUL_MAT_ID
                 }

tools/server/server.cpp

@@ -3812,7 +3812,7 @@ struct server_context {
                             if (slot.n_past > 0 && slot.n_past < (int) slot.prompt.tokens.size()) {
                                 const auto pos_min = llama_memory_seq_pos_min(llama_get_memory(ctx), slot.id);
                                 if (pos_min == -1) {
-                                    SLT_ERR(slot, "n_past = %d, cache_tokens.size() = %d, seq_id = %d, pos_min = %d\n", slot.n_past, (int) slot.prompt.tokens.size(), slot.id, pos_min);
+                                    SLT_ERR(slot, "n_past = %d, slot.prompt.tokens.size() = %d, seq_id = %d, pos_min = %d\n", slot.n_past, (int) slot.prompt.tokens.size(), slot.id, pos_min);
                                     GGML_ABORT("pos_min == -1, but n_past > 0 - should not happen: https://github.com/ggml-org/llama.cpp/pull/13833#discussion_r2116181237");
                                 }
 
@@ -3860,7 +3860,7 @@ struct server_context {
                                 }
 
                                 if (pos_min > pos_min_thold) {
-                                    SLT_WRN(slot, "n_past = %d, cache_tokens.size() = %d, seq_id = %d, pos_min = %d, n_swa = %d\n", slot.n_past, (int) slot.prompt.tokens.size(), slot.id, pos_min, n_swa);
+                                    SLT_WRN(slot, "n_past = %d, slot.prompt.tokens.size() = %d, seq_id = %d, pos_min = %d, n_swa = %d\n", slot.n_past, (int) slot.prompt.tokens.size(), slot.id, pos_min, n_swa);
 
                                     // search for a context checkpoint
                                     const auto it = std::find_if(
@@ -4028,7 +4028,7 @@ struct server_context {
                         }
                     }
 
-                    // SLT_INF(slot, "new cache_tokens: %s\n", slot.cache_tokens.str().c_str());
+                    // SLT_INF(slot, "new slot.prompt.tokens: %s\n", slot.slot.prompt.tokens.str().c_str());
 
                     SLT_INF(slot, "prompt processing progress, n_past = %d, n_tokens = %d, progress = %f\n", slot.n_past, batch.n_tokens, (float) slot.n_past / slot.n_prompt_tokens());
 

tools/server/utils.hpp

@@ -1237,9 +1237,10 @@ public:
             // allowed to resize      ^                    ^
             // disallowed to resize          ^      ^             ^
             if (n > 0) {
-                llama_token last_token = tokens[n - 1];
                 // make sure we never remove tokens in the middle of an image
-                if (last_token == LLAMA_TOKEN_NULL) {
+                // note that the case where we keep a full image at the end is allowed:
+                //   tokens[n - 1] == LLAMA_TOKEN_NULL && tokens[n] != LLAMA_TOKEN_NULL
+                if (tokens[n - 1] == LLAMA_TOKEN_NULL && tokens[n] == LLAMA_TOKEN_NULL) {
                     find_chunk(n - 1); // will throw an error if the token is not begin-of-chunk
                 }
             }