whispercpp 1.3.4 → 1.3.5

data/ext/sources/examples/talk-llama/llama-memory-recurrent.cpp

@@ -7,6 +7,7 @@
 
 #include <algorithm>
 #include <cassert>
+#include <cstring>
 #include <limits>
 #include <map>
 #include <stdexcept>
@@ -32,8 +33,15 @@ llama_memory_recurrent::llama_memory_recurrent(
     cells.clear();
     cells.resize(mem_size);
 
+    // define a comparator for the buft -> ctx map to ensure that the order is well-defined:
+    struct ggml_backend_buft_comparator {
+        bool operator()(const ggml_backend_buffer_type_t & lhs, const ggml_backend_buffer_type_t & rhs) const {
+            return strcmp(ggml_backend_buft_name(lhs), ggml_backend_buft_name(rhs)) < 0;
+        }
+    };
+    std::map<ggml_backend_buffer_type_t, ggml_context_ptr, ggml_backend_buft_comparator> ctx_map;
+
     // create a context for each buffer type
-    std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;
     auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * {
         auto it = ctx_map.find(buft);
         if (it == ctx_map.end()) {
@@ -48,13 +56,12 @@ llama_memory_recurrent::llama_memory_recurrent(
                 return nullptr;
             }
 
-            ctx_map[buft] = ctx;
-            ctxs.emplace_back(ctx);
+            ctx_map.emplace(buft, ctx);
 
             return ctx;
         }
 
-        return it->second;
+        return it->second.get();
     };
 
     r_l.resize(n_layer);
@@ -93,17 +100,14 @@ llama_memory_recurrent::llama_memory_recurrent(
     }
 
     // allocate tensors and initialize the buffers to avoid NaNs in the padding
-    for (auto it : ctx_map) {
-        auto * buft = it.first;
-        auto * ctx  = it.second;
-
-        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
+    for (auto & [buft, ctx] : ctx_map) {
+        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx.get(), buft);
         if (!buf) {
             throw std::runtime_error("failed to allocate buffer for rs cache");
         }
         ggml_backend_buffer_clear(buf, 0);
         LLAMA_LOG_INFO("%s: %10s RS buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf), ggml_backend_buffer_get_size(buf)/1024.0/1024.0);
-        bufs.emplace_back(buf);
+        ctxs_bufs.emplace_back(std::move(ctx), buf);
     }
 
     {
@@ -129,13 +133,14 @@ void llama_memory_recurrent::clear(bool data) {
     used = 0;
 
     if (data) {
-        for (auto & buf : bufs) {
+        for (auto & [_, buf] : ctxs_bufs) {
             ggml_backend_buffer_clear(buf.get(), 0);
         }
     }
 }
 
 bool llama_memory_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
+    //printf("[DEBUG] calling llama_memory_recurrent::seq_rm` with `seq_id=%d, p0=%d, p1=%d`\n", seq_id, p0, p1);
     uint32_t new_head = size;
 
     if (p0 < 0) {
@@ -146,7 +151,8 @@ bool llama_memory_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
         p1 = std::numeric_limits<llama_pos>::max();
     }
 
-    // models like Mamba or RWKV can't have a state partially erased
+    // models like Mamba or RWKV can't have a state partially erased at the end
+    // of the sequence because their state isn't preserved for previous tokens
     if (seq_id >= (int64_t) size) {
         // could be fatal
         return false;
@@ -155,8 +161,9 @@ bool llama_memory_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
         int32_t & tail_id = cells[seq_id].tail;
         if (tail_id >= 0) {
             const auto & cell = cells[tail_id];
-            // partial intersection is invalid
-            if ((0 < p0 && p0 <= cell.pos) || (0 < p1 && p1 <= cell.pos)) {
+            // partial intersection is invalid if it includes the final pos
+            if (0 < p0 && p0 <= cell.pos && p1 > cell.pos) {
+                //printf("[DEBUG] inside `llama_memory_recurrent::seq_rm`: partial intersection is invalid, so returning false\n");
                 return false;
             }
             // invalidate tails which will be cleared
@@ -167,6 +174,7 @@ bool llama_memory_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
     } else {
         // seq_id is negative, then the range should include everything or nothing
         if (p0 != p1 && (p0 != 0 || p1 != std::numeric_limits<llama_pos>::max())) {
+            //printf("[DEBUG] inside `llama_memory_recurrent::seq_rm`: `seq_id` is negative, so returning false\n");
             return false;
         }
     }
@@ -361,8 +369,8 @@ llama_pos llama_memory_recurrent::seq_pos_max(llama_seq_id seq_id) const {
 
 std::map<ggml_backend_buffer_type_t, size_t> llama_memory_recurrent::memory_breakdown() const {
     std::map<ggml_backend_buffer_type_t, size_t> ret;
-    for (const ggml_backend_buffer_ptr & buf_ptr : bufs) {
-        ret[ggml_backend_buffer_get_type(buf_ptr.get())] += ggml_backend_buffer_get_size(buf_ptr.get());
+    for (const auto & [_, buf] : ctxs_bufs) {
+        ret[ggml_backend_buffer_get_type(buf.get())] += ggml_backend_buffer_get_size(buf.get());
     }
     return ret;
 }
@@ -379,7 +387,9 @@ llama_memory_context_ptr llama_memory_recurrent::init_batch(llama_batch_allocr &
                 // if all tokens are output, split by sequence
                 ubatch = balloc.split_seq(n_ubatch);
             } else {
-                ubatch = balloc.split_equal(n_ubatch, false);
+                // TODO: non-sequential equal split can be done if using unified KV cache
+                //       for simplicity, we always use sequential equal split for now
+                ubatch = balloc.split_equal(n_ubatch, true);
             }
 
             if (ubatch.n_tokens == 0) {
@@ -657,7 +667,7 @@ bool llama_memory_recurrent::get_can_shift() const {
 
 size_t llama_memory_recurrent::total_size() const {
     size_t size = 0;
-    for (const auto & buf : bufs) {
+    for (const auto & [_, buf] : ctxs_bufs) {
         size += ggml_backend_buffer_get_size(buf.get());
     }
 
@@ -856,9 +866,12 @@ void llama_memory_recurrent::state_write_data(llama_io_write_i & io, const std::
 bool llama_memory_recurrent::state_read_meta(llama_io_read_i & io, uint32_t cell_count, llama_seq_id dest_seq_id) {
     if (dest_seq_id != -1) {
         // single sequence
-
         seq_rm(dest_seq_id, -1, -1);
 
+        if (cell_count == 0) {
+            return true;
+        }
+
         llama_batch_allocr balloc(hparams.n_pos_per_embd());
 
         llama_ubatch ubatch = balloc.ubatch_reserve(cell_count, 1);

data/ext/sources/examples/talk-llama/llama-memory-recurrent.h

@@ -109,8 +109,8 @@ private:
 
     const uint32_t n_seq_max = 1;
 
-    std::vector<ggml_context_ptr>        ctxs;
-    std::vector<ggml_backend_buffer_ptr> bufs;
+    // ggml contexts for the KV cache along with the allocated backend buffers:
+    std::vector<std::pair<ggml_context_ptr, ggml_backend_buffer_ptr>> ctxs_bufs;
 
     size_t total_size() const;
 

data/ext/sources/examples/talk-llama/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)));
@@ -109,7 +110,7 @@ struct llama_file::impl {
         }
     }
 
-    void read_raw(void * ptr, size_t len) const {
+    void read_raw(void * ptr, size_t len) {
         size_t bytes_read = 0;
         while (bytes_read < len) {
             size_t chunk_size = std::min<size_t>(len - bytes_read, 64*1024*1024);
@@ -126,7 +127,7 @@ struct llama_file::impl {
         }
     }
 
-    uint32_t read_u32() const {
+    uint32_t read_u32() {
         uint32_t val;
         read_raw(&val, sizeof(val));
         return val;
@@ -153,16 +154,55 @@ struct llama_file::impl {
         write_raw(&val, sizeof(val));
     }
 
+    bool has_direct_io() const {
+        return true;
+    }
+
     ~impl() {
         if (fp) {
             std::fclose(fp);
         }
     }
 #else
-    impl(const char * fname, const char * mode) {
-        fp = ggml_fopen(fname, mode);
+    impl(const char * fname, const char * mode, [[maybe_unused]] const bool use_direct_io = false) : fname(fname) {
+#ifdef __linux__
+        // Try unbuffered I/O for read only
+        if (use_direct_io && std::strcmp(mode, "rb") == 0) {
+            if (init_fd()) {
+                return;
+            }
+            LLAMA_LOG_WARN("Failed to open file '%s' with error: %s. Falling back to buffered I/O",
+                           fname, strerror(errno));
+        }
+#endif
+        init_fp(mode);
+    }
+
+#ifdef __linux__
+    bool init_fd() {
+        fd = open(fname.c_str(), 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 true;
+        }
+        return false;
+    }
+#endif
+
+    void init_fp(const char * mode) {
+        fp = ggml_fopen(fname.c_str(), mode);
         if (fp == NULL) {
-            throw std::runtime_error(format("failed to open %s: %s", fname, strerror(errno)));
+            throw std::runtime_error(format("failed to open %s: %s", fname.c_str(), strerror(errno)));
         }
         seek(0, SEEK_END);
         size = tell();
@@ -170,46 +210,118 @@ 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)));
         }
     }
 
-    void read_raw(void * ptr, size_t len) const {
+    void read_raw_unsafe(void * ptr, size_t len) {
         if (len == 0) {
             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 {
+            size_t bytes_read = 0;
+            while (bytes_read < len) {
+                const size_t to_read = len - bytes_read;
+                ssize_t ret = ::read(fd, reinterpret_cast<char *>(ptr) + bytes_read, to_read);
+
+                if (ret == -1) {
+                    if (errno == EINTR) {
+                        continue;  // Interrupted by signal, retry
+                    }
+                    // Fallback to std::fread in case the DMA controller cannot access the buffer
+                    if (errno == EFAULT) {
+                        auto curr_off = tell();
+                        close(fd);
+                        fd = -1;
+                        alignment = 1;
+                        init_fp("rb");
+                        seek(curr_off, SEEK_SET);
+                        read_raw_unsafe(ptr, len);
+                        return;
+                    }
+                    throw std::runtime_error(format("read error: %s", strerror(errno)));
+                }
+                if (ret == 0) {
+                    // EOF: allow if this read was only pulling alignment padding past file end
+                    off_t pos = lseek(fd, 0, SEEK_CUR);
+                    if (pos != -1 && (size_t) pos == size) {
+                        std::memset(reinterpret_cast<char *>(ptr) + bytes_read, 0, len - bytes_read);
+                        return;
+                    }
+                    throw std::runtime_error("unexpectedly reached end of file");
+                }
+
+                bytes_read += (size_t) ret;
+            }
         }
-        if (ret != 1) {
-            throw std::runtime_error("unexpectedly reached end of file");
+    }
+
+    void read_aligned_chunk(void * dest, size_t size) {
+        size_t offset = tell();
+        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_unsafe(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 {
+    void read_raw(void * ptr, size_t len) {
+        if (has_direct_io()) {
+            read_aligned_chunk(ptr, len);
+        } else {
+            read_raw_unsafe(ptr, len);
+        }
+    }
+
+    uint32_t read_u32() {
         uint32_t ret;
         read_raw(&ret, sizeof(ret));
         return ret;
@@ -230,23 +342,41 @@ struct llama_file::impl {
         write_raw(&val, sizeof(val));
     }
 
+    bool has_direct_io() const {
+        return fd != -1 && alignment > 1;
+    }
+
     ~impl() {
-        if (fp) {
+        if (fd != -1) {
+            close(fd);
+        } else {
             std::fclose(fp);
         }
     }
+    int fd = -1;
+    std::string fname;
 #endif
 
-    FILE * fp;
-    size_t size;
+    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(); }
+bool llama_file::has_direct_io() const { return pimpl->has_direct_io(); }
+
 int llama_file::file_id() const {
 #ifdef _WIN32
     return _fileno(pimpl->fp);
@@ -260,9 +390,14 @@ 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(void * ptr, size_t len) { pimpl->read_raw(ptr, len); }
+#ifdef _WIN32
+void llama_file::read_raw_unsafe(void * ptr, size_t len) { pimpl->read_raw(ptr, len); }
+#else
+void llama_file::read_raw_unsafe(void * ptr, size_t len) { pimpl->read_raw_unsafe(ptr, len); }
+#endif
 
-uint32_t llama_file::read_u32() const { return pimpl->read_u32(); }
+uint32_t llama_file::read_u32() { return pimpl->read_u32(); }
 
 void llama_file::write_raw(const void * ptr, size_t len) const { pimpl->write_raw(ptr, len); }
 void llama_file::write_u32(uint32_t val) const { pimpl->write_u32(val); }
@@ -485,7 +620,7 @@ struct llama_mlock::impl {
         if (suggest && getrlimit(RLIMIT_MEMLOCK, &lock_limit)) {
             suggest = false;
         }
-        if (suggest && (lock_limit.rlim_max > lock_limit.rlim_cur + size)) {
+        if (suggest && ((uint64_t)lock_limit.rlim_max > (uint64_t)lock_limit.rlim_cur + size)) {
             suggest = false;
         }
 #endif

data/ext/sources/examples/talk-llama/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;
@@ -23,12 +24,16 @@ struct llama_file {
 
     void seek(size_t offset, int whence) const;
 
-    void read_raw(void * ptr, size_t len) const;
-    uint32_t read_u32() const;
+    void read_raw(void * ptr, size_t len);
+    void read_raw_unsafe(void * ptr, size_t len);
+    void read_aligned_chunk(void * dest, size_t size);
+    uint32_t read_u32();
 
     void write_raw(const void * ptr, size_t len) const;
     void write_u32(uint32_t val) const;
 
+    size_t read_alignment() const;
+    bool has_direct_io() const;
 private:
     struct impl;
     std::unique_ptr<impl> pimpl;

data/ext/sources/examples/talk-llama/llama-model-loader.cpp

@@ -462,15 +462,42 @@ namespace GGUFMeta {
         return get_key_or_arr(llm_kv(kid), result, n, required);
     }
 
+    bool llama_model_loader::get_key_or_arr(enum llm_kv kid, uint32_t & result, bool required) {
+        const std::string key = llm_kv(kid);
+
+        const int id = gguf_find_key(meta.get(), key.c_str());
+
+        if (id < 0) {
+            if (required) {
+                throw std::runtime_error(format("key not found in model: %s", key.c_str()));
+            }
+            return false;
+        }
+
+        // throw and error if type is an array
+        if (gguf_get_kv_type(meta.get(), id) == GGUF_TYPE_ARRAY) {
+            if (required) {
+                throw std::runtime_error(format("expected scalar, found array for key: %s", key.c_str()));
+            }
+            return false;
+        }
+
+        return get_key(key, result, required);
+    }
+
     // TODO: this is not very clever - figure out something better
     template bool llama_model_loader::get_key_or_arr<std::array<int, 4>>(enum llm_kv kid, std::array<int, 4> & result, uint32_t n, bool required);
     template bool llama_model_loader::get_key_or_arr<std::array<uint32_t, 512>>(enum llm_kv kid, std::array<uint32_t, 512> & result, uint32_t n, bool required);
+    template bool llama_model_loader::get_key_or_arr<std::array<float, 512>>(enum llm_kv kid, std::array<float, 512> & result, uint32_t n, bool required);
+
 
 llama_model_loader::llama_model_loader(
         const std::string & fname,
         std::vector<std::string> & splits,
         bool use_mmap,
+        bool use_direct_io,
         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;
@@ -501,9 +528,17 @@ 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_direct_io));
     contexts.emplace_back(ctx);
 
+    use_direct_io = use_direct_io && files.back()->has_direct_io();
+
+    // Disable mmap in case Direct I/O is enabled and available
+    if (use_direct_io && use_mmap) {
+        use_mmap = false;
+        LLAMA_LOG_WARN("%s: direct I/O is enabled, disabling mmap\n", __func__);
+    }
+
     // Save tensors data offset of the main file.
     // For subsidiary files, `meta` tensor data offset must not be used,
     // so we build a unified tensors index for weights.
@@ -569,7 +604,7 @@ 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_direct_io));
             contexts.emplace_back(ctx);
 
             // Save tensors data offset info of the shard.
@@ -713,7 +748,9 @@ llama_model_loader::llama_model_loader(
     }
 
     this->use_mmap = use_mmap;
+    this->use_direct_io = use_direct_io;
     this->check_tensors = check_tensors;
+    this->no_alloc = no_alloc;
 }
 
 std::string llama_model_loader::get_arch_name() const {
@@ -931,7 +968,15 @@ 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;
@@ -981,6 +1026,7 @@ 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));
@@ -1062,6 +1108,7 @@ 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);
@@ -1073,19 +1120,54 @@ 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_unsafe(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;
                     }