docs(readme): document GCN mask_opt and fix Vulkan serving guidance

Add the flash-attn mask_opt change (auto-on for GCN large head sizes,
lossless, +12% pp @ 32k on the RX 580) and correct the earlier pinning
advice: pinning helps in isolated llama-bench but fails in a long-running
server on RADV, so --no-mmap is the serving path. Add a recommended
Polaris serving command.

Assisted-by: Claude

README.md

@@ -12,6 +12,77 @@
 
 LLM inference in C/C++
 
+## ⚡ This fork — Fable's MoE-offload prefill optimizations
+
+Two **opt-in** optimizations for large MoE models whose experts are offloaded to system RAM
+(`--n-cpu-moe`), found and implemented by Fable. Both are **off by default**, toggled via
+environment variables, and produce **token-identical** output to mainline.
+
+| Env var | What it does |
+| --- | --- |
+| `GGML_CUDA_REGISTER_HOST=1` | Page-locks (pins) the mmap'd CPU expert weights so host->device copies go straight over DMA instead of through the driver's hidden bounce buffer (~6-7 -> ~20 GB/s). Works on CUDA and Vulkan (also honored as `GGML_VK_REGISTER_HOST`). Note: it is a presence check, so `=0` still enables it. |
+| `GGML_SCHED_PREFETCH_EXPERTS=1` | Prefetches each layer's experts on a second stream, so the weight uploads overlap compute instead of stalling the GPU. **CUDA only** - on the Vulkan backend the second backend instance shares one device queue, giving no overlap, so this regresses (see Vulkan note below). Leave it off on Vulkan. |
+
+### Benchmark
+
+Measured on an **RTX 3060 12GB** with **Qwen3.6-35B-A3B** (`--n-cpu-moe 26`), prompt-processing at 2048 (`MODEL` = path to your `.gguf`):
+
+```bash
+# baseline (patches off):
+./build/bin/llama-bench -m MODEL -ngl 99 -ncmoe 26 -p 2048 -n 0 -r 5 -b 2048 -ub 2048
+
+# patched (both optimizations on):
+GGML_CUDA_REGISTER_HOST=1 GGML_SCHED_PREFETCH_EXPERTS=1 \
+./build/bin/llama-bench -m MODEL -ngl 99 -ncmoe 26 -p 2048 -n 0 -r 5 -b 2048 -ub 2048
+```
+
+Result: **~1143 → ~1880 t/s** prefill (**+64%**) — same GPU, same settings, token-identical.
+
+Branches: [`fable5/host-register`](https://github.com/thecodacus/llama.cpp/tree/fable5/host-register) (pinning only) · [`fable5/prefetch-experts`](https://github.com/thecodacus/llama.cpp/tree/fable5/prefetch-experts) (both — this branch).
+
+### Vulkan (older AMD, e.g. RX 580 / Polaris)
+
+On the Vulkan backend the CUDA-oriented flags above behave differently, and this fork adds a
+Polaris-specific flash-attention fix. Findings on an **RX 580 8GB** (Polaris / GCN, PCIe 3.0 x16,
+no fp16, no matrix cores) with **Qwen3.5-35B-A3B Q4_K_M**, `-b 2048 -ub 2048`:
+
+- **Flash-attention `mask_opt` is now enabled for GCN large head sizes (this fork's own change).**
+  Upstream disables it on GCN; it is a **lossless** win in high-context prefill - it skips
+  fully-masked causal blocks and the per-block mask add on fully-visible ones, which is real work on
+  a card whose attention is compute-bound (no matrix cores). Auto-on, no flag. On Qwen3.5-35B
+  (head_dim 256): pp2048 **+8% @ 16k, +12% @ 32k**, growing with depth; perplexity bit-identical.
+- **For a long-running server, load with `--no-mmap`, not pinning.** `GGML_CUDA_REGISTER_HOST=1`
+  (pinning) gives ~+17% in an isolated `llama-bench` run, but in a server the RADV host-pointer
+  import fails and the fallback pre-stage buffer allocation fails for large / co-resident models, so
+  it silently reverts to slow staging (and can trip warnings/OOM). `--no-mmap` (weights in RAM) is
+  both faster and clean there. Pinning is still fine for one-off `llama-bench` numbers.
+- **`GGML_SCHED_PREFETCH_EXPERTS=1` regresses - do not use it** on Vulkan (its second backend shares
+  one device queue, so uploads never overlap compute).
+- **`-b 2048 -ub 2048` is the biggest prefill lever** (the default `-ub 512` roughly halves pp).
+- **Tune `--n-cpu-moe` to context length.** Keep some expert layers resident in spare VRAM for short
+  prompts (e.g. `ncmoe 28` on the 35B, ~+5% over all-host); at long context the KV cache needs that
+  VRAM, so raise it (`ncmoe 40`, all experts on host). Keep flash attention on (`-fa 1`).
+
+Prefill throughput (isolated `llama-bench`, pinned unless noted):
+
+| Config | pp2048 (t/s) |
+| --- | ---: |
+| baseline, unpinned, `ncmoe 40` | ~252 |
+| pinned, `ncmoe 40` | ~294 |
+| pinned, `ncmoe 28` | ~308 |
+| server default (`--no-mmap`, `ncmoe 40`) | ~285 |
+| + `mask_opt`, @ 32k context | **+12%** |
+
+**Recommended RX 580 / Polaris serving command** (per model):
+
+```bash
+llama-server -hf <repo>:<quant> --no-mmap -ngl 99 --n-cpu-moe 40 -b 2048 -ub 2048 -fa 1
+```
+
+Lower `--n-cpu-moe` (e.g. 28) if the model plus your context budget leave spare VRAM; keep it high
+for long-context / agentic use. At long context the bottleneck is attention compute (GPU-bound), so
+`mask_opt` (above) is where the remaining prefill wins come from, not the MoE-transfer path.
+
 ## Recent API changes
 
 - [Changelog for `libllama` API](https://github.com/ggml-org/llama.cpp/issues/9289)

ggml/src/ggml-backend.cpp

@@ -761,6 +761,10 @@ static bool ggml_is_view_op(enum ggml_op op) {
 #define GGML_SCHED_MAX_COPIES 4
 #endif
 
+#ifndef GGML_SCHED_MAX_PREFETCH_SLOTS
+#define GGML_SCHED_MAX_PREFETCH_SLOTS 8
+#endif
+
 struct ggml_backend_sched_split {
     int backend_id;
     int i_start;
@@ -818,6 +822,19 @@ struct ggml_backend_sched {
 
     bool op_offload;
 
+    // full-tensor prefetch of offloaded MUL_MAT_ID weights (GGML_SCHED_PREFETCH_EXPERTS)
+    // with a large batch virtually every expert is used, so the routing ids are not worth
+    // waiting for; uploads run through a second backend instance on the same device so
+    // they overlap compute, alternating between two staging slots
+    bool prefetch_experts;
+    ggml_backend_t prefetch_backend;
+    int prefetch_n_slots;
+    ggml_backend_buffer_t prefetch_slots[GGML_SCHED_MAX_PREFETCH_SLOTS];
+    ggml_backend_event_t prefetch_ready[GGML_SCHED_MAX_PREFETCH_SLOTS];
+    ggml_backend_event_t prefetch_free[GGML_SCHED_MAX_PREFETCH_SLOTS];
+    bool prefetch_used[GGML_SCHED_MAX_PREFETCH_SLOTS];
+    int prefetch_cur;
+
     int debug;
 
     // used for debugging graph reallocations [GGML_SCHED_DEBUG_REALLOC]
@@ -1538,6 +1555,94 @@ static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
     return true;
 }
 
+static void ggml_backend_sched_prefetch_disable(ggml_backend_sched_t sched, ggml_backend_t split_backend) {
+    sched->prefetch_experts = false;
+    if (sched->prefetch_backend) {
+        ggml_backend_synchronize(split_backend);
+        ggml_backend_synchronize(sched->prefetch_backend);
+    }
+    for (int i = 0; i < sched->prefetch_n_slots; i++) {
+        ggml_backend_buffer_free(sched->prefetch_slots[i]);
+        sched->prefetch_slots[i] = NULL;
+        sched->prefetch_used[i] = false;
+    }
+}
+
+// slots are sized once for the largest offloaded expert tensor in the current graph so
+// that they never need to grow mid-eval
+static size_t ggml_backend_sched_prefetch_max_size(ggml_backend_sched_t sched) {
+    size_t max_size = 0;
+    for (int split_id = 0; split_id < sched->n_splits; split_id++) {
+        struct ggml_backend_sched_split * split = &sched->splits[split_id];
+        if (split->graph.n_nodes == 0 || split->graph.nodes[0]->op != GGML_OP_MUL_MAT_ID) {
+            continue;
+        }
+        for (int input_id = 0; input_id < split->n_inputs; input_id++) {
+            const ggml_tensor * input = split->inputs[input_id];
+            if (input->buffer &&
+                ggml_backend_buffer_get_usage(input->buffer) == GGML_BACKEND_BUFFER_USAGE_WEIGHTS &&
+                ggml_backend_buffer_is_host(input->buffer)) {
+                max_size = std::max(max_size, ggml_nbytes(input));
+            }
+        }
+    }
+    return max_size;
+}
+
+static bool ggml_backend_sched_prefetch_init(ggml_backend_sched_t sched, ggml_backend_t split_backend, size_t size) {
+    if (sched->prefetch_backend == NULL) {
+        ggml_backend_dev_t dev = split_backend->device;
+        ggml_backend_dev_props props;
+        ggml_backend_dev_get_props(dev, &props);
+        if (!props.caps.async || !props.caps.events) {
+            sched->prefetch_experts = false;
+            return false;
+        }
+        sched->prefetch_backend = ggml_backend_dev_init(dev, NULL);
+        if (sched->prefetch_backend == NULL) {
+            sched->prefetch_experts = false;
+            return false;
+        }
+        for (int i = 0; i < sched->prefetch_n_slots; i++) {
+            sched->prefetch_ready[i] = ggml_backend_event_new(dev);
+            sched->prefetch_free[i]  = ggml_backend_event_new(dev);
+            if (sched->prefetch_ready[i] == NULL || sched->prefetch_free[i] == NULL) {
+                sched->prefetch_experts = false;
+                return false;
+            }
+        }
+    }
+
+    size = std::max(size, ggml_backend_sched_prefetch_max_size(sched));
+
+    ggml_backend_buffer_type_t buft = ggml_backend_get_default_buffer_type(split_backend);
+    for (int i = 0; i < sched->prefetch_n_slots; i++) {
+        if (sched->prefetch_slots[i] == NULL || ggml_backend_buffer_get_size(sched->prefetch_slots[i]) < size) {
+            // allocate before freeing so a failure leaves the old slot intact
+            ggml_backend_buffer_t new_buf = ggml_backend_buft_alloc_buffer(buft, size);
+            if (new_buf == NULL) {
+                // overlap needs at least 2 slots, otherwise run with what fits
+                if (i >= 2 && sched->prefetch_slots[0] != NULL &&
+                    ggml_backend_buffer_get_size(sched->prefetch_slots[0]) >= size) {
+                    sched->prefetch_n_slots = i;
+                    sched->prefetch_cur = 0;
+                    return true;
+                }
+                ggml_backend_sched_prefetch_disable(sched, split_backend);
+                return false;
+            }
+            if (sched->prefetch_slots[i] != NULL) {
+                ggml_backend_synchronize(split_backend);
+                ggml_backend_synchronize(sched->prefetch_backend);
+                ggml_backend_buffer_free(sched->prefetch_slots[i]);
+            }
+            sched->prefetch_slots[i] = new_buf;
+            sched->prefetch_used[i] = false;
+        }
+    }
+    return true;
+}
+
 static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t sched) {
     GGML_ASSERT(sched);
     struct ggml_backend_sched_split * splits = sched->splits;
@@ -1550,6 +1655,10 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
         struct ggml_backend_sched_split * split = &splits[split_id];
         int split_backend_id = split->backend_id;
         ggml_backend_t split_backend = sched->backends[split_backend_id];
+        int split_prefetch_slot = -1;
+        ggml_tensor * prefetch_input_cpy = NULL;
+        ggml_backend_buffer_t prefetch_saved_buffer = NULL;
+        void * prefetch_saved_data = NULL;
 
         // copy the input tensors to the split backend
         for (int input_id = 0; input_id < split->n_inputs; input_id++) {
@@ -1566,6 +1675,41 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
                 }
                 ggml_backend_tensor_copy(input, input_cpy);
             } else {
+                // with a large batch virtually every expert is used, so instead of waiting
+                // for the routing ids, upload the full tensor through the prefetch backend
+                // and let the copy overlap compute of the previous split
+                if (sched->prefetch_experts && !sched->callback_eval && split_prefetch_slot == -1 && split->graph.n_nodes > 0) {
+                    ggml_tensor * node = split->graph.nodes[0];
+                    if (ggml_backend_buffer_get_usage(input->buffer) == GGML_BACKEND_BUFFER_USAGE_WEIGHTS &&
+                        ggml_backend_buffer_is_host(input->buffer) &&
+                        node->op == GGML_OP_MUL_MAT_ID && node->src[0] == input_cpy) {
+                        const ggml_tensor * ids = node->src[2];
+                        const int64_t n_expert = input->ne[2];
+                        if (ids->ne[0]*ids->ne[1] >= 2*n_expert &&
+                            ggml_backend_sched_prefetch_init(sched, split_backend, ggml_nbytes(input))) {
+                            const int slot = sched->prefetch_cur;
+                            sched->prefetch_cur = (sched->prefetch_cur + 1) % sched->prefetch_n_slots;
+                            // wait for the previous user of this slot to finish computing
+                            if (sched->prefetch_used[slot]) {
+                                ggml_backend_event_wait(sched->prefetch_backend, sched->prefetch_free[slot]);
+                            }
+                            // point the staging copy at the slot only for the duration of
+                            // this split, so a fallback to the regular path on a later
+                            // eval can never see a dangling slot pointer
+                            prefetch_input_cpy    = input_cpy;
+                            prefetch_saved_buffer = input_cpy->buffer;
+                            prefetch_saved_data   = input_cpy->data;
+                            input_cpy->buffer = sched->prefetch_slots[slot];
+                            input_cpy->data   = ggml_backend_buffer_get_base(sched->prefetch_slots[slot]);
+                            ggml_backend_tensor_set_async(sched->prefetch_backend, input_cpy, input->data, 0, ggml_nbytes(input));
+                            ggml_backend_event_record(sched->prefetch_ready[slot], sched->prefetch_backend);
+                            ggml_backend_event_wait(split_backend, sched->prefetch_ready[slot]);
+                            split_prefetch_slot = slot;
+                            continue;
+                        }
+                    }
+                }
+
                 // wait for the split backend to finish using the input before overwriting it
                 if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
                     ggml_backend_event_wait(split_backend, sched->events[split_backend_id][sched->cur_copy]);
@@ -1676,6 +1820,13 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
 
         if (!sched->callback_eval) {
             enum ggml_status ec = ggml_backend_graph_compute_async(split_backend, &split->graph);
+            if (split_prefetch_slot != -1) {
+                // the kernels have captured the slot address at launch, safe to restore
+                ggml_backend_event_record(sched->prefetch_free[split_prefetch_slot], split_backend);
+                sched->prefetch_used[split_prefetch_slot] = true;
+                prefetch_input_cpy->buffer = prefetch_saved_buffer;
+                prefetch_input_cpy->data   = prefetch_saved_data;
+            }
             if (ec != GGML_STATUS_SUCCESS) {
                 return ec;
             }
@@ -1788,6 +1939,15 @@ ggml_backend_sched_t ggml_backend_sched_new(
     sched->galloc = ggml_gallocr_new_n(sched->bufts, n_backends);
     sched->op_offload = op_offload;
 
+    // GGML_SCHED_PREFETCH_EXPERTS=1 enables the default slot count, higher values set it
+    // directly; more slots let uploads run further ahead of compute at the cost of one
+    // max-sized expert tensor of device memory per slot
+    const char * GGML_SCHED_PREFETCH_EXPERTS = getenv("GGML_SCHED_PREFETCH_EXPERTS");
+    const int prefetch_n_slots = GGML_SCHED_PREFETCH_EXPERTS ? atoi(GGML_SCHED_PREFETCH_EXPERTS) : 0;
+    sched->prefetch_experts = op_offload && prefetch_n_slots > 0;
+    // default of 3 covers the gate/up/down expert tensors of one MoE layer
+    sched->prefetch_n_slots = prefetch_n_slots <= 1 ? 3 : std::min(prefetch_n_slots, GGML_SCHED_MAX_PREFETCH_SLOTS);
+
     ggml_backend_sched_reset(sched);
 
     return sched;
@@ -1802,6 +1962,16 @@ void ggml_backend_sched_free(ggml_backend_sched_t sched) {
             ggml_backend_event_free(sched->events[b][c]);
         }
     }
+    if (sched->prefetch_backend) {
+        ggml_backend_synchronize(sched->prefetch_backend);
+        // the slot count may have been reduced after a failed allocation, free everything
+        for (int i = 0; i < GGML_SCHED_MAX_PREFETCH_SLOTS; i++) {
+            ggml_backend_event_free(sched->prefetch_ready[i]);
+            ggml_backend_event_free(sched->prefetch_free[i]);
+            ggml_backend_buffer_free(sched->prefetch_slots[i]);
+        }
+        ggml_backend_free(sched->prefetch_backend);
+    }
     ggml_gallocr_free(sched->galloc);
     ggml_free(sched->ctx);
     ggml_hash_set_free(&sched->hash_set);
@@ -1906,6 +2076,9 @@ void ggml_backend_sched_synchronize(ggml_backend_sched_t sched) {
     for (int i = 0; i < sched->n_backends; i++) {
         ggml_backend_synchronize(sched->backends[i]);
     }
+    if (sched->prefetch_backend) {
+        ggml_backend_synchronize(sched->prefetch_backend);
+    }
     if (!sched->is_alloc) {
         // if the graph is not already allocated, always use copy 0 after a synchronization
         // this ensures that during generation the same copy is used every time,

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -3283,6 +3283,7 @@ static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, const std
             import_info.setPNext(&mem_flags_info);
             buf->device_memory = device->device.allocateMemory({ size, memory_type_idx, &import_info });
         } catch (const vk::SystemError& e) {
+            GGML_LOG_WARN("ggml_vulkan: host pointer memory import failed (%s)\n", e.what());
         }
     } else {
         for (auto it = req_flags_list.begin(); it != req_flags_list.end(); it++) {
@@ -8128,25 +8129,32 @@ static bool ggml_vk_buffer_write_2d_async(vk_context subctx, vk_buffer& dst, siz
     ggml_vk_host_get(dst->device, src, buf, buf_offset);
 
     if (buf != nullptr) {
-        // Memory is pinned, use as staging buffer
-        std::vector<vk::BufferCopy> slices(1);
-        if (width == spitch && width == dpitch) {
-            // Only do single write if stride is equal
-            slices[0].srcOffset = buf_offset;
-            slices[0].dstOffset = offset;
-            slices[0].size = width * height;
-        } else {
-            slices.resize(height);
-            for (size_t i = 0; i < height; i++) {
-                slices[i].srcOffset = buf_offset + i * spitch;
-                slices[i].dstOffset = offset + i * dpitch;
-                slices[i].size = width;
+        // extent of the read in pinned source memory; guard against tensors that
+        // straddle a pinned-chunk boundary (they fall back to staging below)
+        size_t src_extent = (width == spitch) ? (size_t) width * height
+                                              : (height > 0 ? (height - 1) * spitch + width : 0);
+        if (buf_offset + src_extent <= buf->size) {
+            // Memory is pinned, use as staging buffer
+            std::vector<vk::BufferCopy> slices(1);
+            if (width == spitch && width == dpitch) {
+                // Only do single write if stride is equal
+                slices[0].srcOffset = buf_offset;
+                slices[0].dstOffset = offset;
+                slices[0].size = width * height;
+            } else {
+                slices.resize(height);
+                for (size_t i = 0; i < height; i++) {
+                    slices[i].srcOffset = buf_offset + i * spitch;
+                    slices[i].dstOffset = offset + i * dpitch;
+                    slices[i].size = width;
+                }
             }
-        }
 
-        ggml_vk_sync_buffers(nullptr, subctx);
-        subctx->s->buffer->buf.copyBuffer(buf->buffer, dst->buffer, slices);
-        return true;
+            ggml_vk_sync_buffers(nullptr, subctx);
+            subctx->s->buffer->buf.copyBuffer(buf->buffer, dst->buffer, slices);
+            return true;
+        }
+        // straddles a chunk boundary: fall through to staging
     }
     VK_LOG_DEBUG("STAGING");
 
@@ -10584,8 +10592,10 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     }
 
     // Only use mask opt when the mask is fairly large. This hasn't been tuned extensively.
+    // GCN with large head size (>= 256) benefits in high-context prefill (skipping the
+    // per-block mask add on fully-visible blocks dominates), so enable it there too.
     bool use_mask_opt = mask && nem1 >= 32 && nem0 * nem1 > 32768 && nem0 >= tuning_params.block_cols * 16
-                        && (ctx->device->architecture != vk_device_architecture::AMD_GCN || HSK > 256 || HSV > 256);
+                        && (ctx->device->architecture != vk_device_architecture::AMD_GCN || HSK >= 256 || HSV >= 256);
     vk_fa_pipeline_state fa_pipeline_state = get_fa_pipeline_state(ctx->device, tuning_params, HSK, HSV, aligned, f32acc,
                                                                    mask != nullptr, use_mask_opt, logit_softcap != 0, k->type, v->type);
 
@@ -12512,9 +12522,108 @@ static void ggml_vk_opt_step_sgd(ggml_backend_vk_context * ctx, vk_context& subc
     ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, src2, nullptr, dst, GGML_OP_OPT_STEP_SGD, { (uint32_t)n, 0, 0.0f, 0.0f, 0.0f, 0.0f });
 }
 
+// Fast path for concat along dim 0 where one source is stored "transposed"
+// (nb[1] == type_size, dim1 innermost) and the other source + dst are
+// contiguous along dim 0. The generic concat shader reads the transposed
+// source with a catastrophic uncoalesced stride; here we instead copy the
+// contiguous source with copy.comp and transpose the other source into the
+// matching dst sub-region with the tiled copy_transpose shader (shared-memory
+// transpose, coalesced read+write). Mirrors the precedent in
+// ggml_vk_cpy_to_contiguous: direct dispatch with custom push constants.
+static void ggml_vk_concat_transpose_fastpath(ggml_backend_vk_context * ctx, vk_context& subctx,
+                                              const ggml_tensor * ctg, const ggml_tensor * trp,
+                                              ggml_tensor * dst, uint32_t off_ctg, uint32_t off_trp) {
+    const uint32_t ts = ggml_type_size(dst->type);
+
+    vk_pipeline pipeline_cpy = (ts == 4) ? ctx->device->pipeline_cpy_f32_f32
+                                          : ctx->device->pipeline_cpy_f16_f16;
+    vk_pipeline pipeline_trp = (ts == 4) ? ctx->device->pipeline_cpy_transpose_32
+                                          : ctx->device->pipeline_cpy_transpose_16;
+
+    ggml_pipeline_request_descriptor_sets(ctx, pipeline_cpy, 1);
+    ggml_pipeline_request_descriptor_sets(ctx, pipeline_trp, 1);
+
+    vk_subbuffer ctg_buf = ggml_vk_tensor_subbuffer(ctx, ctg, true);
+    vk_subbuffer trp_buf = ggml_vk_tensor_subbuffer(ctx, trp, true);
+    vk_subbuffer dst_buf = ggml_vk_tensor_subbuffer(ctx, dst, true);
+
+    const uint32_t a_misalign_ctg = get_misalign_bytes(ctx, ctg) / ts;
+    const uint32_t a_misalign_trp = get_misalign_bytes(ctx, trp) / ts;
+    const uint32_t d_misalign     = get_misalign_bytes(ctx, dst) / ts;
+
+    // Dispatch A: contiguous copy of `ctg` into dst[off_ctg : off_ctg + ctg->ne[0], :]
+    if (ctg->ne[0] > 0) {
+        const uint32_t ne_ctg = (uint32_t) ggml_nelements(ctg);
+        vk_op_unary_push_constants pc = vk_op_unary_push_constants_init(ctg, dst, ne_ctg);
+        pc.ne10 = (uint32_t) ctg->ne[0];                 // only ctg's columns in dst
+        pc.misalign_offsets = (a_misalign_ctg << 16) | (d_misalign + off_ctg);
+        init_pushconst_fastdiv(pc);
+        std::array<uint32_t, 3> el = ne_ctg > 262144 ? std::array<uint32_t,3>{512, 512, CEIL_DIV(ne_ctg, 262144)}
+                                      : ne_ctg > 512  ? std::array<uint32_t,3>{512, CEIL_DIV(ne_ctg, 512), 1}
+                                                      : std::array<uint32_t,3>{ne_ctg, 1, 1};
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline_cpy, { ctg_buf, dst_buf }, pc, el);
+    }
+
+    // Dispatch B: tiled transpose of `trp` into dst[off_trp : off_trp + trp->ne[0], :]
+    if (trp->ne[0] > 0) {
+        vk_op_unary_push_constants pc = vk_op_unary_push_constants_init(trp, dst, ggml_nelements(trp));
+        pc.ne10 = (uint32_t) trp->ne[0];                 // dst bound = trp columns (NOT dst->ne[0])
+        pc.misalign_offsets = (a_misalign_trp << 16) | (d_misalign + off_trp);
+        init_pushconst_fastdiv(pc);
+        std::array<uint32_t, 3> el = {
+            (uint32_t) CEIL_DIV(trp->ne[0], 32),
+            (uint32_t) CEIL_DIV(trp->ne[1], 32),
+            (uint32_t) (trp->ne[2] * trp->ne[3]),
+        };
+        el[0] = std::min(el[0], (uint32_t) ctx->device->properties.limits.maxComputeWorkGroupCount[0]);
+        el[1] = std::min(el[1], (uint32_t) ctx->device->properties.limits.maxComputeWorkGroupCount[1]);
+        el[2] = std::min(el[2], (uint32_t) ctx->device->properties.limits.maxComputeWorkGroupCount[2]);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline_trp, { trp_buf, dst_buf }, pc, el);
+    }
+
+    ggml_vk_sync_buffers(ctx, subctx);
+}
+
 static void ggml_vk_concat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     int * op_params = (int *)dst->op_params;
 
+    // Fast path: concat along dim 0 with one source "transposed" (nb[1]==type_size,
+    // dim1 innermost) and the other source + dst contiguous along dim 0. The generic
+    // shader reads the transposed source uncoalesced (~5.6ms on RX 580 vs ~130us for
+    // a coalesced copy); here we instead use a tiled shared-memory transpose.
+    auto src_transposed_2d = [&](const ggml_tensor * s) {
+        const uint32_t ts = ggml_type_size(s->type);
+        return s->nb[1] == ts                                              // dim1 innermost
+            && s->nb[0] == s->ne[1] * ts                                   // consistent 2D transpose
+            && s->ne[2] == 1 && s->ne[3] == 1;
+    };
+    const uint32_t dst_ts = ggml_type_size(dst->type);
+    const bool dim0_ok   = (op_params[0] == 0) && (dst->nb[0] == dst_ts);
+    const bool types_ok  = (dst_ts == 4 || dst_ts == 2)
+                        && (src0->type == src1->type && src0->type == dst->type)
+                        && (ggml_blck_size(dst->type) == 1);
+    const bool shapes_ok = (src0->ne[1] == src1->ne[1] && src1->ne[1] == dst->ne[1])
+                        && (src0->ne[2] == src1->ne[2] && src0->ne[2] == dst->ne[2])
+                        && (src0->ne[3] == src1->ne[3] && src0->ne[3] == dst->ne[3])
+                        && (src0->ne[0] + src1->ne[0] == dst->ne[0]);
+    // doffset is 16-bit for unary shaders; guard against overflow
+    const bool off_fits  = ((get_misalign_bytes(ctx, dst)/dst_ts + dst->ne[0])) < 0xFFFFu;
+
+    const bool s1_trans = dim0_ok && types_ok && shapes_ok && off_fits
+                        && src_transposed_2d(src1) && ggml_is_contiguous(src0);
+    const bool s0_trans = dim0_ok && types_ok && shapes_ok && off_fits
+                        && src_transposed_2d(src0) && ggml_is_contiguous(src1);
+
+    if (s1_trans || s0_trans) {
+        const ggml_tensor * ctg = s1_trans ? src0 : src1;     // contiguous source
+        const ggml_tensor * trp = s1_trans ? src1 : src0;     // transposed source
+        // dst offsets (in elements) where each source region begins
+        const uint32_t off_ctg = s1_trans ? 0u                       : (uint32_t) src1->ne[0];
+        const uint32_t off_trp = s1_trans ? (uint32_t) src0->ne[0]   : 0u;
+        ggml_vk_concat_transpose_fastpath(ctx, subctx, ctg, trp, dst, off_ctg, off_trp);
+        return;
+    }
+
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t src1_type_size = ggml_type_size(src1->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
@@ -16578,7 +16687,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
         std::fill(ctx->query_nodes.begin(), ctx->query_nodes.end(), nullptr);
         std::fill(ctx->query_node_idx.begin(), ctx->query_node_idx.end(), 0);
 
-        GGML_ASSERT(ctx->compute_ctx.expired());
+        // compute_ctx may hold async host uploads recorded before the graph; append the timestamp after them
         compute_ctx = ggml_vk_get_compute_ctx(ctx);
         ctx->query_idx = 0;
         compute_ctx->s->buffer->buf.writeTimestamp(vk::PipelineStageFlagBits::eAllCommands, ctx->query_pool, ctx->query_idx++);
@@ -18243,11 +18352,112 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
     return devices[device];
 }
 
+// Import an mmap-backed host region as a Vulkan pinned buffer via
+// VK_EXT_external_memory_host so H2D uploads DMA straight from system RAM
+// instead of bouncing through the staging buffer + host memcpy. Mirrors the
+// GGML_CUDA_REGISTER_HOST path; populates device->pinned_memory, which the
+// existing pinned fast path in ggml_vk_buffer_write_2d_async looks up.
+//
+// A single Vulkan buffer cannot cover a whole multi-GB mmap (it is capped at
+// device->max_buffer_size), so the region is imported in page-aligned chunks.
+// ggml_vk_host_get resolves a tensor pointer to the chunk that contains it,
+// and ggml_vk_buffer_write_2d_async falls back to staging for any tensor that
+// straddles a chunk boundary, so correctness is preserved.
+static bool ggml_backend_vk_register_host_buffer(void * buffer, size_t size) {
+    if (getenv("GGML_CUDA_REGISTER_HOST") == nullptr && getenv("GGML_VK_REGISTER_HOST") == nullptr) {
+        return false;
+    }
+    if (size == 0) {
+        return false;
+    }
+
+    bool success = false;
+    for (size_t i = 0; i < GGML_VK_MAX_DEVICES; i++) {
+        vk_device& device = vk_instance.devices[i];
+        if (!device || !device->external_memory_host || device->max_buffer_size == 0) {
+            continue;
+        }
+
+        const size_t align = device->min_imported_host_pointer_alignment;
+        size_t chunk = device->max_buffer_size & ~(align - 1);
+        if (chunk == 0) {
+            continue;
+        }
+
+        uint8_t * p = static_cast<uint8_t *>(buffer);
+        size_t remaining = size;
+        bool dev_success = false;
+        bool import_ok = true;   // flips to false once VK_EXT_external_memory_host import fails
+        while (remaining > 0) {
+            size_t cur = std::min(remaining, chunk);
+            vk_buffer buf;
+            if (import_ok) {
+                buf = ggml_vk_buffer_from_host_ptr(device, p, cur);
+            }
+            if (!buf || !buf->buffer) {
+                // Fallback for drivers that can't import file-backed mmap pages (e.g. RADV):
+                // make a one-time copy of the region into a host-visible Vulkan buffer. The GPU
+                // then DMAs straight from it every eval at full PCIe bandwidth, instead of paying
+                // a slow single-threaded pageable memcpy into the staging buffer each time.
+                import_ok = false;
+                buf = ggml_vk_create_buffer_check(device, cur,
+                    vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached,
+                    vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
+                if (buf && buf->buffer && buf->ptr) {
+                    memcpy(buf->ptr, p, cur);
+                } else {
+                    break;
+                }
+            }
+            {
+                std::lock_guard<std::shared_mutex> guard(device->pinned_memory_mutex);
+                device->pinned_memory.emplace_back(p, cur, buf);
+            }
+            dev_success = true;
+            p += cur;
+            remaining -= cur;
+        }
+        if (dev_success) {
+            success = true;
+        }
+    }
+    return success;
+}
+
+static void ggml_backend_vk_unregister_host_buffer(void * buffer) {
+    for (size_t i = 0; i < GGML_VK_MAX_DEVICES; i++) {
+        vk_device& device = vk_instance.devices[i];
+        if (!device) {
+            continue;
+        }
+        std::lock_guard<std::shared_mutex> guard(device->pinned_memory_mutex);
+        for (auto it = device->pinned_memory.begin(); it != device->pinned_memory.end(); ++it) {
+            if (std::get<0>(*it) == buffer) {
+                vk_buffer buf = std::get<2>(*it);
+                device->pinned_memory.erase(it);
+                ggml_vk_destroy_buffer(buf);
+                break;
+            }
+        }
+    }
+}
+
+static void * ggml_backend_vk_reg_get_proc_address(ggml_backend_reg_t reg, const char * name) {
+    GGML_UNUSED(reg);
+    if (strcmp(name, "ggml_backend_register_host_buffer") == 0) {
+        return (void *) ggml_backend_vk_register_host_buffer;
+    }
+    if (strcmp(name, "ggml_backend_unregister_host_buffer") == 0) {
+        return (void *) ggml_backend_vk_unregister_host_buffer;
+    }
+    return nullptr;
+}
+
 static const struct ggml_backend_reg_i ggml_backend_vk_reg_i = {
     /* .get_name         = */ ggml_backend_vk_reg_get_name,
     /* .get_device_count = */ ggml_backend_vk_reg_get_device_count,
     /* .get_device       = */ ggml_backend_vk_reg_get_device,
-    /* .get_proc_address = */ NULL,
+    /* .get_proc_address = */ ggml_backend_vk_reg_get_proc_address,
 };
 
 ggml_backend_reg_t ggml_backend_vk_reg() {

src/llama-mmap.cpp

@@ -618,13 +618,47 @@ struct llama_mmap::impl {
 };
 
 llama_mmap::llama_mmap(struct llama_file * file, size_t prefetch, bool numa) : pimpl(std::make_unique<impl>(file, prefetch, numa)) {}
-llama_mmap::~llama_mmap() = default;
+
+llama_mmap::~llama_mmap() {
+    // unpin before the pages are unmapped by the impl destructor
+    if (host_reg_addr && host_unreg_fn) {
+        host_unreg_fn(host_reg_addr);
+    }
+}
 
 size_t llama_mmap::size() const { return pimpl->size; }
 void * llama_mmap::addr() const { return pimpl->addr; }
 
 void llama_mmap::unmap_fragment(size_t first, size_t last) { pimpl->unmap_fragment(first, last); }
 
+size_t llama_mmap::register_host(size_t first, size_t last, bool (*reg_fn)(void *, size_t), void (*unreg_fn)(void *)) {
+#ifdef _POSIX_MAPPED_FILES
+    if (host_reg_addr || !reg_fn || !unreg_fn || last <= first) {
+        return 0;
+    }
+
+    // expand outward to the page boundaries retained by unmap_fragment
+    const size_t page_size = sysconf(_SC_PAGESIZE);
+    first = first & ~(page_size - 1);
+    last  = (last + page_size - 1) & ~(page_size - 1);
+
+    void * reg_addr = (uint8_t *) pimpl->addr + first;
+    if (!reg_fn(reg_addr, last - first)) {
+        return 0;
+    }
+
+    host_reg_addr  = reg_addr;
+    host_unreg_fn  = unreg_fn;
+    return last - first;
+#else
+    GGML_UNUSED(first);
+    GGML_UNUSED(last);
+    GGML_UNUSED(reg_fn);
+    GGML_UNUSED(unreg_fn);
+    return 0;
+#endif
+}
+
 #if defined(_POSIX_MEMLOCK_RANGE) || defined(_WIN32)
 const bool llama_mmap::SUPPORTED  = true;
 #else

src/llama-mmap.h

@@ -50,11 +50,19 @@ struct llama_mmap {
 
     void unmap_fragment(size_t first, size_t last);
 
+    // pin the pages backing [first, last) with a backend allocator for faster H2D copies,
+    // unpinned in the destructor before the pages are unmapped
+    // returns the number of bytes registered, 0 on failure
+    size_t register_host(size_t first, size_t last, bool (*reg_fn)(void *, size_t), void (*unreg_fn)(void *));
+
     static const bool SUPPORTED;
 
 private:
     struct impl;
     std::unique_ptr<impl> pimpl;
+
+    void * host_reg_addr = nullptr;
+    void (*host_unreg_fn)(void *) = nullptr;
 };
 
 struct llama_mlock {

src/llama-model-loader.cpp

@@ -1677,6 +1677,15 @@ bool llama_model_loader::load_all_data(
     if (size_done >= size_data) {
         // unmap offloaded tensors and metadata
         if (use_mmap) {
+            // pin the pages backing the weights kept in system memory for faster H2D copies
+            bool (*reg_fn)(void *, size_t) = nullptr;
+            void (*unreg_fn)(void *) = nullptr;
+            for (size_t i = 0; i < ggml_backend_dev_count() && !reg_fn; i++) {
+                ggml_backend_reg_t reg = ggml_backend_dev_backend_reg(ggml_backend_dev_get(i));
+                reg_fn   = (bool (*)(void *, size_t)) ggml_backend_reg_get_proc_address(reg, "ggml_backend_register_host_buffer");
+                unreg_fn = (void (*)(void *))         ggml_backend_reg_get_proc_address(reg, "ggml_backend_unregister_host_buffer");
+            }
+
             for (uint32_t idx = 0; idx < mappings.size(); idx++) {
                 const auto & mmap_used = mmaps_used.at(idx);
                 auto & mapping = mappings.at(idx);
@@ -1684,6 +1693,13 @@ bool llama_model_loader::load_all_data(
                 if (mmap_used.second != 0) {
                     mapping->unmap_fragment(mmap_used.second, mapping->size());
                 }
+                if (mmap_used.second > mmap_used.first) {
+                    size_t n_registered = mapping->register_host(mmap_used.first, mmap_used.second, reg_fn, unreg_fn);
+                    if (n_registered > 0) {
+                        LLAMA_LOG_INFO("%s: pinned %.2f MiB of mapped model memory for faster H2D transfers\n",
+                                __func__, n_registered / 1024.0 / 1024.0);
+                    }
+                }
             }
         }
         if (progress_callback) {