llama-cpp-pydist 0.18.0__py3-none-any.whl → 0.19.0__py3-none-any.whl

vendor_llama_cpp_pydist/llama.cpp/src/llama-kv-cache.cpp

@@ -175,7 +175,15 @@ llama_kv_cache::llama_kv_cache(
 
     // allocate tensors and initialize the buffers to avoid NaNs in the padding
     for (auto & [buft, ctx] : ctx_map) {
-        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx.get(), buft);
+        ggml_backend_buffer_t buf;
+        if (model.hparams.no_alloc) {
+            buf = ggml_backend_buft_alloc_buffer(buft, /*size =*/ 0); // dummy buffer
+            for (ggml_tensor * t = ggml_get_first_tensor(ctx.get()); t != nullptr; t = ggml_get_next_tensor(ctx.get(), t)) {
+                t->buffer = buf; // set dummy buffer for KV cache so that the backend scheduler won't try to allocate it
+            }
+        } else {
+            buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx.get(), buft); // real buffer
+        }
         if (!buf) {
             throw std::runtime_error("failed to allocate buffer for kv cache");
         }
@@ -482,9 +490,18 @@ llama_pos llama_kv_cache::seq_pos_max(llama_seq_id seq_id) const {
 
 std::map<ggml_backend_buffer_type_t, size_t> llama_kv_cache::memory_breakdown() const {
     std::map<ggml_backend_buffer_type_t, size_t> ret;
-    for (const auto & [_, buf] : ctxs_bufs) {
-        ret[ggml_backend_buffer_get_type(buf.get())] += ggml_backend_buffer_get_size(buf.get());
+    for (const auto & [ctx, buf] : ctxs_bufs) {
+        ggml_backend_buffer_type_t buft = ggml_backend_buffer_get_type(buf.get());
+
+        if (hparams.no_alloc) {
+            GGML_ASSERT(ggml_backend_buffer_get_base(buf.get()) == nullptr);
+            ret[buft] += ggml_backend_alloc_ctx_tensors_from_buft_size(ctx.get(), buft);
+        } else {
+            // GGML_ASSERT(ggml_backend_buffer_get_base(buf.get()) != nullptr); // multi_buffer does not have a defined base
+            ret[buft] += ggml_backend_buffer_get_size(buf.get());
+        }
     }
+
     return ret;
 }
 
@@ -1372,7 +1389,7 @@ ggml_tensor * llama_kv_cache::build_rope_shift(
     const auto & yarn_ext_factor  = cparams.yarn_ext_factor;
     const auto & yarn_beta_fast   = cparams.yarn_beta_fast;
     const auto & yarn_beta_slow   = cparams.yarn_beta_slow;
-    const auto & yarn_attn_factor = llama_hparams::yarn_attn_factor_adjust(cparams.yarn_attn_factor, cparams.rope_freq_scale, cparams.yarn_ext_factor);
+    const auto & yarn_attn_factor = cparams.yarn_attn_factor;
 
     const auto & n_rot     = hparams.n_rot;
     const auto & rope_type = hparams.rope_type == LLAMA_ROPE_TYPE_MROPE || hparams.rope_type == LLAMA_ROPE_TYPE_IMROPE
@@ -1544,9 +1561,11 @@ void llama_kv_cache::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama
 
         const uint32_t strm = seq_id == -1 ? s : seq_to_stream[seq_id];
 
+        slot_info sinfo;
+
         bool res = true;
-        res = res && state_read_meta(io, strm, cell_count, seq_id);
-        res = res && state_read_data(io, strm, cell_count);
+        res = res && state_read_meta(io, strm, cell_count, sinfo, seq_id);
+        res = res && state_read_data(io, strm, cell_count, sinfo);
 
         if (!res) {
             if (seq_id == -1) {
@@ -1685,7 +1704,7 @@ void llama_kv_cache::state_write_data(llama_io_write_i & io, const cell_ranges_t
     }
 }
 
-bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32_t cell_count, llama_seq_id dest_seq_id) {
+bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32_t cell_count, slot_info & sinfo, llama_seq_id dest_seq_id) {
     auto & cells = v_cells[strm];
     auto & head  = v_heads[strm];
 
@@ -1722,7 +1741,7 @@ bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32
             ubatch.seq_id[i]   = &dest_seq_id;
         }
 
-        const auto sinfo = find_slot(ubatch, true);
+        sinfo = find_slot(ubatch, false);
         if (sinfo.empty()) {
             LLAMA_LOG_ERROR("%s: failed to find available cells in kv cache\n", __func__);
             return false;
@@ -1732,20 +1751,16 @@ bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32
         //       see: https://github.com/ggml-org/llama.cpp/pull/16825#issuecomment-3460868350
         apply_ubatch(sinfo, ubatch);
 
-        const auto head_cur = sinfo.head();
+        LLAMA_LOG_DEBUG("%s: cell_count = %d, dest_seq_id = %d\n", __func__, cell_count, dest_seq_id);
 
-        // keep the head at the old position because we will read the KV data into it in state_read_data()
-        head = head_cur;
-
-        LLAMA_LOG_DEBUG("%s: head_cur = %d, head = %d, cell_count = %d, dest_seq_id = %d\n", __func__, head_cur, head, cell_count, dest_seq_id);
-
-        // DEBUG CHECK: head_cur should be our first cell, head_cur + cell_count - 1 should be our last cell (verify seq_id and pos values)
-        // Assume that this is one contiguous block of cells
-        GGML_ASSERT(head_cur + cell_count <= cells.size());
-        GGML_ASSERT(cells.pos_get(head_cur)                  == ubatch.pos[0]);
-        GGML_ASSERT(cells.pos_get(head_cur + cell_count - 1) == ubatch.pos[cell_count - 1]);
-        GGML_ASSERT(cells.seq_has(head_cur,                  dest_seq_id));
-        GGML_ASSERT(cells.seq_has(head_cur + cell_count - 1, dest_seq_id));
+        // DEBUG CHECK: verify that all cells were allocated and have correct seq_id and pos values
+        GGML_ASSERT(sinfo.n_stream() == 1);
+        GGML_ASSERT(sinfo.idxs[0].size() == cell_count);
+        for (uint32_t i = 0; i < cell_count; ++i) {
+            const uint32_t idx = sinfo.idxs[0][i];
+            GGML_ASSERT(cells.pos_get(idx) == ubatch.pos[i]);
+            GGML_ASSERT(cells.seq_has(idx, dest_seq_id));
+        }
     } else {
         // whole KV cache restore
 
@@ -1778,15 +1793,24 @@ bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32
             }
         }
 
+        // Create contiguous slot_info for whole cache restore
+        sinfo.s0 = strm;
+        sinfo.s1 = strm;
+        sinfo.resize(1);
+        sinfo.strm[0] = strm;
+        sinfo.idxs[0].resize(cell_count);
+        for (uint32_t i = 0; i < cell_count; ++i) {
+            sinfo.idxs[0][i] = i;
+        }
+
         head = 0;
     }
 
     return true;
 }
 
-bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32_t cell_count) {
+bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32_t cell_count, const slot_info & sinfo) {
     auto & cells = v_cells[strm];
-    auto & head  = v_heads[strm];
 
     uint32_t v_trans;
     uint32_t n_layer;
@@ -1836,8 +1860,17 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
         }
 
         if (cell_count) {
-            // Read and set the keys for the whole cell range
-            ggml_backend_tensor_set(k, io.read(cell_count * k_size_row), head * k_size_row, cell_count * k_size_row);
+            if (sinfo.is_contiguous()) {
+                // Fast path: contiguous cells, single memcpy
+                ggml_backend_tensor_set(k, io.read(cell_count * k_size_row), sinfo.head() * k_size_row, cell_count * k_size_row);
+            } else {
+                // Slow path: scatter to non-contiguous positions
+                const void * src = io.read(cell_count * k_size_row);
+                for (uint32_t i = 0; i < cell_count; ++i) {
+                    const size_t dst_offset = sinfo.idxs[0][i] * k_size_row;
+                    ggml_backend_tensor_set(k, (const char*)src + i * k_size_row, dst_offset, k_size_row);
+                }
+            }
         }
     }
 
@@ -1868,8 +1901,17 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
             }
 
             if (cell_count) {
-                // Read and set the values for the whole cell range
-                ggml_backend_tensor_set(v, io.read(cell_count * v_size_row), head * v_size_row, cell_count * v_size_row);
+                if (sinfo.is_contiguous()) {
+                    // Fast path: contiguous cells, single memcpy
+                    ggml_backend_tensor_set(v, io.read(cell_count * v_size_row), sinfo.head() * v_size_row, cell_count * v_size_row);
+                } else {
+                    // Slow path: scatter to non-contiguous positions
+                    const void * src = io.read(cell_count * v_size_row);
+                    for (uint32_t i = 0; i < cell_count; ++i) {
+                        const size_t dst_offset = sinfo.idxs[0][i] * v_size_row;
+                        ggml_backend_tensor_set(v, (const char*)src + i * v_size_row, dst_offset, v_size_row);
+                    }
+                }
             }
         }
     } else {
@@ -1908,10 +1950,22 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
             }
 
             if (cell_count) {
-                // For each row in the transposed matrix, read the values for the whole cell range
-                for (uint32_t j = 0; j < n_embd_v_gqa; ++j) {
-                    const size_t dst_offset = (head + j * cells.size()) * v_size_el;
-                    ggml_backend_tensor_set(v, io.read(cell_count * v_size_el), dst_offset, cell_count * v_size_el);
+                if (sinfo.is_contiguous()) {
+                    // Fast path: contiguous cells
+                    const uint32_t h = sinfo.head();
+                    for (uint32_t j = 0; j < n_embd_v_gqa; ++j) {
+                        const size_t dst_offset = (h + j * cells.size()) * v_size_el;
+                        ggml_backend_tensor_set(v, io.read(cell_count * v_size_el), dst_offset, cell_count * v_size_el);
+                    }
+                } else {
+                    // Slow path: scatter to non-contiguous positions
+                    for (uint32_t j = 0; j < n_embd_v_gqa; ++j) {
+                        const void * src = io.read(cell_count * v_size_el);
+                        for (uint32_t i = 0; i < cell_count; ++i) {
+                            const size_t dst_offset = (sinfo.idxs[0][i] + j * cells.size()) * v_size_el;
+                            ggml_backend_tensor_set(v, (const char*)src + i * v_size_el, dst_offset, v_size_el);
+                        }
+                    }
                 }
             }
         }

vendor_llama_cpp_pydist/llama.cpp/src/llama-kv-cache.h

@@ -72,6 +72,23 @@ public:
         void clear() {
             idxs.clear();
         }
+
+        // check if indices are contiguous starting from head()
+        bool is_contiguous() const {
+            if (idxs.empty() || idxs[0].empty()) {
+                return true;
+            }
+            if (idxs.size() > 1) {
+                return false;
+            }
+            const uint32_t h = idxs[0][0];
+            for (size_t i = 0; i < idxs[0].size(); ++i) {
+                if (idxs[0][i] != h + i) {
+                    return false;
+                }
+            }
+            return true;
+        }
     };
 
     using slot_info_vec_t = std::vector<slot_info>;
@@ -264,8 +281,8 @@ private:
     void state_write_meta(llama_io_write_i & io, const cell_ranges_t & cr, llama_seq_id seq_id = -1) const;
     void state_write_data(llama_io_write_i & io, const cell_ranges_t & cr) const;
 
-    bool state_read_meta(llama_io_read_i & io, uint32_t strm, uint32_t cell_count, llama_seq_id dest_seq_id = -1);
-    bool state_read_data(llama_io_read_i & io, uint32_t strm, uint32_t cell_count);
+    bool state_read_meta(llama_io_read_i & io, uint32_t strm, uint32_t cell_count,       slot_info & sinfo, llama_seq_id dest_seq_id = -1);
+    bool state_read_data(llama_io_read_i & io, uint32_t strm, uint32_t cell_count, const slot_info & sinfo);
 };
 
 class llama_kv_cache_context : public llama_memory_context_i {

vendor_llama_cpp_pydist/llama.cpp/src/llama-memory-hybrid.cpp

@@ -222,7 +222,7 @@ llama_memory_hybrid_context::llama_memory_hybrid_context(
     ubatches(std::move(ubatches)),
     // note: here we copy the ubatches. not sure if this is ideal
     ctx_attn(new llama_kv_cache_context(mem->get_mem_attn(), std::move(sinfos_attn), this->ubatches)),
-    ctx_recr(new llama_memory_recurrent_context(mem->get_mem_recr(),                        this->ubatches)),
+    ctx_recr(new llama_memory_recurrent_context(mem->get_mem_recr(), this->ubatches)),
     status(llama_memory_status_combine(ctx_attn->get_status(), ctx_recr->get_status())) {
 }
 

vendor_llama_cpp_pydist/llama.cpp/src/llama-mmap.cpp

@@ -13,9 +13,10 @@
 #ifdef __has_include
     #if __has_include(<unistd.h>)
         #include <unistd.h>
+        #include <fcntl.h>
+        #include <sys/stat.h>
         #if defined(_POSIX_MAPPED_FILES)
             #include <sys/mman.h>
-            #include <fcntl.h>
         #endif
         #if defined(_POSIX_MEMLOCK_RANGE)
             #include <sys/resource.h>
@@ -74,7 +75,7 @@ struct llama_file::impl {
         return ret;
     }
 
-    impl(const char * fname, const char * mode) {
+    impl(const char * fname, const char * mode, [[maybe_unused]] const bool use_direct_io = false) {
         fp = ggml_fopen(fname, mode);
         if (fp == NULL) {
             throw std::runtime_error(format("failed to open %s: %s", fname, strerror(errno)));
@@ -153,13 +154,40 @@ struct llama_file::impl {
         write_raw(&val, sizeof(val));
     }
 
+    void read_aligned_chunk(size_t offset, void * dest, size_t size) const {
+        throw std::runtime_error("DirectIO is not implemented on Windows.");
+    }
+
     ~impl() {
         if (fp) {
             std::fclose(fp);
         }
     }
 #else
-    impl(const char * fname, const char * mode) {
+    impl(const char * fname, const char * mode, [[maybe_unused]] const bool use_direct_io = false) {
+#ifdef __linux__
+        // Try unbuffered I/O for read only
+        if (use_direct_io && std::strcmp(mode, "rb") == 0) {
+            fd = open(fname, O_RDONLY | O_DIRECT);
+
+            if (fd != -1) {
+                struct stat file_stats{};
+                fstat(fd, &file_stats);
+
+                size = file_stats.st_size;
+                alignment = file_stats.st_blksize;
+
+                off_t ret = lseek(fd, 0, SEEK_SET);
+                if (ret == -1) {
+                    throw std::runtime_error(format("seek error: %s", strerror(errno)));
+                }
+                return;
+            }
+
+            LLAMA_LOG_WARN("Failed to open model %s with error: %s. Falling back to buffered I/O",
+                fname, strerror(errno));
+        }
+#endif
         fp = ggml_fopen(fname, mode);
         if (fp == NULL) {
             throw std::runtime_error(format("failed to open %s: %s", fname, strerror(errno)));
@@ -170,27 +198,30 @@ struct llama_file::impl {
     }
 
     size_t tell() const {
-// TODO: this ifdef is never true?
-#ifdef _WIN32
-        __int64 ret = _ftelli64(fp);
-#else
-        long ret = std::ftell(fp);
-#endif
-        if (ret == -1) {
-            throw std::runtime_error(format("ftell error: %s", strerror(errno)));
+        if (fd == -1) {
+            long ret = std::ftell(fp);
+            if (ret == -1) {
+                throw std::runtime_error(format("ftell error: %s", strerror(errno)));
+            }
+
+            return (size_t) ret;
         }
 
-        return (size_t) ret;
+        off_t pos = lseek(fd, 0, SEEK_CUR);
+        if (pos == -1) {
+            throw std::runtime_error(format("lseek error: %s", strerror(errno)));
+        }
+        return (size_t) pos;
     }
 
     void seek(size_t offset, int whence) const {
-// TODO: this ifdef is never true?
-#ifdef _WIN32
-        int ret = _fseeki64(fp, (__int64) offset, whence);
-#else
-        int ret = std::fseek(fp, (long) offset, whence);
-#endif
-        if (ret != 0) {
+        off_t ret = 0;
+        if (fd == -1) {
+            ret = std::fseek(fp, (long) offset, whence);
+        } else {
+            ret = lseek(fd, offset, whence);
+        }
+        if (ret == -1) {
             throw std::runtime_error(format("seek error: %s", strerror(errno)));
         }
     }
@@ -200,13 +231,55 @@ struct llama_file::impl {
             return;
         }
         errno = 0;
-        std::size_t ret = std::fread(ptr, len, 1, fp);
-        if (ferror(fp)) {
-            throw std::runtime_error(format("read error: %s", strerror(errno)));
+        if (fd == -1) {
+            std::size_t ret = std::fread(ptr, len, 1, fp);
+            if (ferror(fp)) {
+                throw std::runtime_error(format("read error: %s", strerror(errno)));
+            }
+            if (ret != 1) {
+                throw std::runtime_error("unexpectedly reached end of file");
+            }
+        } else {
+            bool successful = false;
+            while (!successful) {
+                off_t ret = read(fd, ptr, len);
+
+                if (ret == -1) {
+                    if (errno == EINTR) {
+                        continue;  // Interrupted by signal, retry
+                    }
+                    throw std::runtime_error(format("read error: %s", strerror(errno)));
+                }
+                if (ret == 0) {
+                    throw std::runtime_error("unexpectedly reached end of file");
+                }
+
+                successful = true;
+            }
         }
-        if (ret != 1) {
-            throw std::runtime_error("unexpectedly reached end of file");
+    }
+
+    void read_aligned_chunk(size_t offset, void * dest, size_t size) const {
+        off_t aligned_offset = offset & ~(alignment - 1);
+        off_t offset_from_alignment = offset - aligned_offset;
+        size_t bytes_to_read = (offset_from_alignment + size + alignment - 1) & ~(alignment - 1);
+
+        void * raw_buffer = nullptr;
+        int ret = posix_memalign(&raw_buffer, alignment, bytes_to_read);
+        if (ret != 0) {
+            throw std::runtime_error(format("posix_memalign failed with error %d", ret));
         }
+
+        struct aligned_buffer_deleter {
+            void operator()(void * p) const { free(p); }
+        };
+        std::unique_ptr<void, aligned_buffer_deleter> buffer(raw_buffer);
+
+        seek(aligned_offset, SEEK_SET);
+        read_raw(buffer.get(), bytes_to_read);
+
+        uintptr_t actual_data = reinterpret_cast<uintptr_t>(buffer.get()) + offset_from_alignment;
+        memcpy(dest, reinterpret_cast<void *>(actual_data), size);
     }
 
     uint32_t read_u32() const {
@@ -231,22 +304,43 @@ struct llama_file::impl {
     }
 
     ~impl() {
-        if (fp) {
+        if (fd != -1) {
+            close(fd);
+        } else {
             std::fclose(fp);
         }
     }
+    int fd = -1;
 #endif
 
-    FILE * fp;
-    size_t size;
+    void read_raw_at(void * ptr, size_t len, size_t offset) const {
+        if (alignment != 1) {
+            read_aligned_chunk(offset, ptr, len);
+        } else {
+            seek(offset, SEEK_SET);
+            read_raw(ptr, len);
+        }
+    }
+
+    size_t read_alignment() const {
+        return alignment;
+    }
+
+    size_t alignment = 1;
+
+    FILE * fp{};
+    size_t size{};
 };
 
-llama_file::llama_file(const char * fname, const char * mode) : pimpl(std::make_unique<impl>(fname, mode)) {}
+llama_file::llama_file(const char * fname, const char * mode, const bool use_direct_io) :
+    pimpl(std::make_unique<impl>(fname, mode, use_direct_io)) {}
 llama_file::~llama_file() = default;
 
 size_t llama_file::tell() const { return pimpl->tell(); }
 size_t llama_file::size() const { return pimpl->size; }
 
+size_t llama_file::read_alignment() const { return pimpl->read_alignment(); }
+
 int llama_file::file_id() const {
 #ifdef _WIN32
     return _fileno(pimpl->fp);
@@ -261,6 +355,7 @@ int llama_file::file_id() const {
 
 void llama_file::seek(size_t offset, int whence) const { pimpl->seek(offset, whence); }
 void llama_file::read_raw(void * ptr, size_t len) const { pimpl->read_raw(ptr, len); }
+void llama_file::read_raw_at(void * ptr, size_t len, size_t offset) const { pimpl->read_raw_at(ptr, len, offset); }
 
 uint32_t llama_file::read_u32() const { return pimpl->read_u32(); }
 

vendor_llama_cpp_pydist/llama.cpp/src/llama-mmap.h

@@ -3,6 +3,7 @@
 #include <cstdint>
 #include <memory>
 #include <vector>
+#include <cstdio>
 
 struct llama_file;
 struct llama_mmap;
@@ -13,7 +14,7 @@ using llama_mmaps  = std::vector<std::unique_ptr<llama_mmap>>;
 using llama_mlocks = std::vector<std::unique_ptr<llama_mlock>>;
 
 struct llama_file {
-    llama_file(const char * fname, const char * mode);
+    llama_file(const char * fname, const char * mode, bool use_direct_io = false);
     ~llama_file();
 
     size_t tell() const;
@@ -24,11 +25,14 @@ struct llama_file {
     void seek(size_t offset, int whence) const;
 
     void read_raw(void * ptr, size_t len) const;
+    void read_raw_at(void * ptr, size_t len, size_t offset) const;
+    void read_aligned_chunk(size_t offset, void * dest, size_t size) const;
     uint32_t read_u32() const;
 
     void write_raw(const void * ptr, size_t len) const;
     void write_u32(uint32_t val) const;
 
+    size_t read_alignment() const;
 private:
     struct impl;
     std::unique_ptr<impl> pimpl;

vendor_llama_cpp_pydist/llama.cpp/src/llama-model-loader.cpp

@@ -473,6 +473,7 @@ llama_model_loader::llama_model_loader(
         std::vector<std::string> & splits,
         bool use_mmap,
         bool check_tensors,
+        bool no_alloc,
         const llama_model_kv_override * param_overrides_p,
         const llama_model_tensor_buft_override * param_tensor_buft_overrides_p) {
     int trace = 0;
@@ -503,7 +504,7 @@ llama_model_loader::llama_model_loader(
     get_key(llm_kv(LLM_KV_GENERAL_ARCHITECTURE), arch_name, false);
     llm_kv = LLM_KV(llm_arch_from_string(arch_name));
 
-    files.emplace_back(new llama_file(fname.c_str(), "rb"));
+    files.emplace_back(new llama_file(fname.c_str(), "rb", !use_mmap));
     contexts.emplace_back(ctx);
 
     // Save tensors data offset of the main file.
@@ -571,7 +572,7 @@ llama_model_loader::llama_model_loader(
                 }
             }
 
-            files.emplace_back(new llama_file(fname_split, "rb"));
+            files.emplace_back(new llama_file(fname_split, "rb", !use_mmap));
             contexts.emplace_back(ctx);
 
             // Save tensors data offset info of the shard.
@@ -716,6 +717,7 @@ llama_model_loader::llama_model_loader(
 
     this->use_mmap = use_mmap;
     this->check_tensors = check_tensors;
+    this->no_alloc = no_alloc;
 }
 
 std::string llama_model_loader::get_arch_name() const {
@@ -933,7 +935,15 @@ bool llama_model_loader::load_all_data(
     // 4 staging buffers for async uploads, each sized 1MB seems to be a good default for single NVMe drives.
     // NVMe raid configurations might require more / larger buffers.
     constexpr size_t n_buffers = 4;
-    constexpr size_t buffer_size = 1 * 1024 * 1024; // 1MB
+
+    size_t alignment = 1;
+    for (const auto & file : files) {
+        alignment = std::max(file->read_alignment(), alignment);
+    }
+
+    // Buffer size: balance between memory usage and I/O efficiency
+    // 64MB works well for NVMe drives
+    const size_t buffer_size = alignment != 1 ? 64 * 1024 * 1024 + 2 * alignment : 1 * 1024 * 1024;
 
     std::vector<ggml_backend_buffer_t> host_buffers;
     std::vector<ggml_backend_event_t> events;
@@ -983,6 +993,7 @@ bool llama_model_loader::load_all_data(
         // If the backend is supported, create pinned memory buffers and events for synchronisation.
         for (size_t idx = 0; idx < n_buffers; ++idx) {
             auto * buf = ggml_backend_buft_alloc_buffer(host_buft, buffer_size);
+
             if (!buf) {
                 LLAMA_LOG_DEBUG("%s: failed to allocate host buffer for async uploads for device %s\n", func,
                     ggml_backend_dev_name(dev));
@@ -1064,9 +1075,9 @@ bool llama_model_loader::load_all_data(
             }
         } else {
             const auto & file = files.at(weight->idx);
+
             if (ggml_backend_buffer_is_host(cur->buffer)) {
-                file->seek(weight->offs, SEEK_SET);
-                file->read_raw(cur->data, n_size);
+                file->read_raw_at(cur->data, n_size, weight->offs);
                 if (check_tensors) {
                     validation_result.emplace_back(std::async(std::launch::async, [cur, n_size] {
                         return std::make_pair(cur, ggml_validate_row_data(cur->type, cur->data, n_size));
@@ -1075,26 +1086,60 @@ bool llama_model_loader::load_all_data(
             } else {
                 // If upload_backend is valid load the tensor in chunks to pinned memory and upload the buffers asynchronously to the GPU.
                 if (upload_backend) {
-                    file->seek(weight->offs, SEEK_SET);
+                    size_t offset = weight->offs;
+                    alignment = file->read_alignment();
+                    size_t aligned_offset = offset & ~(alignment - 1);
+                    size_t offset_from_alignment = offset - aligned_offset;
+                    file->seek(aligned_offset, SEEK_SET);
+
+                    // Calculate aligned read boundaries
+                    size_t read_start = aligned_offset;
+                    size_t read_end = (offset + n_size + alignment - 1) & ~(alignment - 1);
 
                     size_t bytes_read = 0;
+                    size_t data_read = 0;  // Actual tensor data copied (excluding padding)
+
+                    while (bytes_read < read_end - read_start) {
+                        size_t read_size = std::min<size_t>(buffer_size, read_end - read_start - bytes_read);
 
-                    while (bytes_read < n_size) {
-                        size_t read_iteration = std::min<size_t>(buffer_size, n_size - bytes_read);
+                        // Align the destination pointer within the pinned buffer
+                        uintptr_t ptr_dest_aligned = (reinterpret_cast<uintptr_t>(host_ptrs[buffer_idx]) + alignment - 1) & ~(alignment - 1);
 
+                        // Wait for previous upload to complete before reusing buffer
                         ggml_backend_event_synchronize(events[buffer_idx]);
-                        file->read_raw(host_ptrs[buffer_idx], read_iteration);
-                        ggml_backend_tensor_set_async(upload_backend, cur, host_ptrs[buffer_idx], bytes_read, read_iteration);
+
+                        // Read aligned chunk from file
+                        file->read_raw(reinterpret_cast<void *>(ptr_dest_aligned), read_size);
+
+                        // Calculate actual data portion (excluding alignment padding)
+                        uintptr_t ptr_data = ptr_dest_aligned;
+                        size_t data_to_copy = read_size;
+
+                        // Skip alignment padding at start of first chunk
+                        if (bytes_read == 0) {
+                            ptr_data += offset_from_alignment;
+                            data_to_copy -= offset_from_alignment;
+                        }
+
+                        // Trim alignment padding at end of last chunk
+                        if (aligned_offset + bytes_read + read_size > offset + n_size) {
+                            data_to_copy -= (read_end - (offset + n_size));
+                        }
+
+                        // Async upload actual data to GPU
+                        ggml_backend_tensor_set_async(upload_backend, cur,
+                                                      reinterpret_cast<void *>(ptr_data), data_read, data_to_copy);
                         ggml_backend_event_record(events[buffer_idx], upload_backend);
 
-                        bytes_read += read_iteration;
+                        data_read += data_to_copy;
+                        bytes_read += read_size;
+
                         ++buffer_idx;
                         buffer_idx %= n_buffers;
                     }
                 } else {
                     read_buf.resize(n_size);
-                    file->seek(weight->offs, SEEK_SET);
-                    file->read_raw(read_buf.data(), n_size);
+                    file->read_raw_at(read_buf.data(), n_size, weight->offs);
                     ggml_backend_tensor_set(cur, read_buf.data(), 0, n_size);
                     if (check_tensors && !ggml_validate_row_data(cur->type, read_buf.data(), n_size)) {
                         throw std::runtime_error(format("tensor '%s' has invalid data", ggml_get_name(cur)));

vendor_llama_cpp_pydist/llama.cpp/src/llama-model-loader.h

@@ -71,6 +71,7 @@ struct llama_model_loader {
 
     bool use_mmap = false;
     bool check_tensors;
+    bool no_alloc;
 
     llama_files files;
     llama_ftype ftype;
@@ -97,6 +98,7 @@ struct llama_model_loader {
         std::vector<std::string> & splits, // optional, only need if the split does not follow naming scheme
         bool use_mmap,
         bool check_tensors,
+        bool no_alloc,
         const llama_model_kv_override * param_overrides_p,
         const llama_model_tensor_buft_override * param_tensor_buft_overrides_p);