Handle missing model in CLI parameters for llama-run (#11399)

The HTTP client in llama-run only prints an error in case the download of
a resource failed. If the model name in the CLI parameter list is missing,
this causes the application to crash.
In order to prevent this, a check for the required model parameter has been
added and errors for resource downloads get propagated to the caller.

Signed-off-by: Michael Engel <mengel@redhat.com>

examples/run/run.cpp

@@ -181,6 +181,10 @@ class Opt {
             }
         }
 
+        if (model_.empty()){
+            return 1;
+        }
+
         return 0;
     }
 
@@ -319,6 +323,10 @@ class HttpClient {
   public:
     int init(const std::string & url, const std::vector<std::string> & headers, const std::string & output_file,
              const bool progress, std::string * response_str = nullptr) {
+        if (std::filesystem::exists(output_file)) {
+            return 0;
+        }
+
         std::string output_file_partial;
         curl = curl_easy_init();
         if (!curl) {
@@ -346,7 +354,11 @@ class HttpClient {
         data.file_size = set_resume_point(output_file_partial);
         set_progress_options(progress, data);
         set_headers(headers);
-        perform(url);
+        CURLcode res = perform(url);
+        if (res != CURLE_OK){
+            printe("Fetching resource '%s' failed: %s\n", url.c_str(), curl_easy_strerror(res));
+            return 1;
+        }
         if (!output_file.empty()) {
             std::filesystem::rename(output_file_partial, output_file);
         }
@@ -411,16 +423,12 @@ class HttpClient {
         }
     }
 
-    void perform(const std::string & url) {
-        CURLcode res;
+    CURLcode perform(const std::string & url) {
         curl_easy_setopt(curl, CURLOPT_URL, url.c_str());
         curl_easy_setopt(curl, CURLOPT_FOLLOWLOCATION, 1L);
         curl_easy_setopt(curl, CURLOPT_DEFAULT_PROTOCOL, "https");
         curl_easy_setopt(curl, CURLOPT_FAILONERROR, 1L);
-        res = curl_easy_perform(curl);
-        if (res != CURLE_OK) {
-            printe("curl_easy_perform() failed: %s\n", curl_easy_strerror(res));
-        }
+        return curl_easy_perform(curl);
     }
 
     static std::string human_readable_time(double seconds) {
@@ -558,13 +566,14 @@ class LlamaData {
         }
 
         sampler = initialize_sampler(opt);
+
         return 0;
     }
 
   private:
 #ifdef LLAMA_USE_CURL
-    int download(const std::string & url, const std::vector<std::string> & headers, const std::string & output_file,
-                 const bool progress, std::string * response_str = nullptr) {
+    int download(const std::string & url, const std::string & output_file, const bool progress,
+                 const std::vector<std::string> & headers = {}, std::string * response_str = nullptr) {
         HttpClient http;
         if (http.init(url, headers, output_file, progress, response_str)) {
             return 1;
@@ -573,48 +582,85 @@ class LlamaData {
         return 0;
     }
 #else
-    int download(const std::string &, const std::vector<std::string> &, const std::string &, const bool,
+    int download(const std::string &, const std::string &, const bool, const std::vector<std::string> & = {},
                  std::string * = nullptr) {
         printe("%s: llama.cpp built without libcurl, downloading from an url not supported.\n", __func__);
+
         return 1;
     }
 #endif
 
-    int huggingface_dl(const std::string & model, const std::vector<std::string> headers, const std::string & bn) {
-        // Find the second occurrence of '/' after protocol string
-        size_t pos = model.find('/');
-        pos        = model.find('/', pos + 1);
-        if (pos == std::string::npos) {
-            return 1;
-        }
-
-        const std::string hfr = model.substr(0, pos);
-        const std::string hff = model.substr(pos + 1);
-        const std::string url = "https://huggingface.co/" + hfr + "/resolve/main/" + hff;
-        return download(url, headers, bn, true);
-    }
-
-    int ollama_dl(std::string & model, const std::vector<std::string> headers, const std::string & bn) {
-        if (model.find('/') == std::string::npos) {
-            model = "library/" + model;
-        }
-
-        std::string model_tag = "latest";
-        size_t      colon_pos = model.find(':');
+    // Helper function to handle model tag extraction and URL construction
+    std::pair<std::string, std::string> extract_model_and_tag(std::string & model, const std::string & base_url) {
+        std::string  model_tag = "latest";
+        const size_t colon_pos = model.find(':');
         if (colon_pos != std::string::npos) {
             model_tag = model.substr(colon_pos + 1);
             model     = model.substr(0, colon_pos);
         }
 
-        std::string manifest_url = "https://registry.ollama.ai/v2/" + model + "/manifests/" + model_tag;
+        std::string url = base_url + model + "/manifests/" + model_tag;
+
+        return { model, url };
+    }
+
+    // Helper function to download and parse the manifest
+    int download_and_parse_manifest(const std::string & url, const std::vector<std::string> & headers,
+                                    nlohmann::json & manifest) {
         std::string manifest_str;
-        const int   ret = download(manifest_url, headers, "", false, &manifest_str);
+        int         ret = download(url, "", false, headers, &manifest_str);
         if (ret) {
             return ret;
         }
 
-        nlohmann::json manifest = nlohmann::json::parse(manifest_str);
-        std::string    layer;
+        manifest = nlohmann::json::parse(manifest_str);
+
+        return 0;
+    }
+
+    int huggingface_dl(std::string & model, const std::string & bn) {
+        // Find the second occurrence of '/' after protocol string
+        size_t pos = model.find('/');
+        pos        = model.find('/', pos + 1);
+        std::string              hfr, hff;
+        std::vector<std::string> headers = { "User-Agent: llama-cpp", "Accept: application/json" };
+        std::string              url;
+
+        if (pos == std::string::npos) {
+            auto [model_name, manifest_url] = extract_model_and_tag(model, "https://huggingface.co/v2/");
+            hfr                             = model_name;
+
+            nlohmann::json manifest;
+            int            ret = download_and_parse_manifest(manifest_url, headers, manifest);
+            if (ret) {
+                return ret;
+            }
+
+            hff = manifest["ggufFile"]["rfilename"];
+        } else {
+            hfr = model.substr(0, pos);
+            hff = model.substr(pos + 1);
+        }
+
+        url = "https://huggingface.co/" + hfr + "/resolve/main/" + hff;
+
+        return download(url, bn, true, headers);
+    }
+
+    int ollama_dl(std::string & model, const std::string & bn) {
+        const std::vector<std::string> headers = { "Accept: application/vnd.docker.distribution.manifest.v2+json" };
+        if (model.find('/') == std::string::npos) {
+            model = "library/" + model;
+        }
+
+        auto [model_name, manifest_url] = extract_model_and_tag(model, "https://registry.ollama.ai/v2/");
+        nlohmann::json manifest;
+        int            ret = download_and_parse_manifest(manifest_url, {}, manifest);
+        if (ret) {
+            return ret;
+        }
+
+        std::string layer;
         for (const auto & l : manifest["layers"]) {
             if (l["mediaType"] == "application/vnd.ollama.image.model") {
                 layer = l["digest"];
@@ -622,8 +668,9 @@ class LlamaData {
             }
         }
 
-        std::string blob_url = "https://registry.ollama.ai/v2/" + model + "/blobs/" + layer;
-        return download(blob_url, headers, bn, true);
+        std::string blob_url = "https://registry.ollama.ai/v2/" + model_name + "/blobs/" + layer;
+
+        return download(blob_url, bn, true, headers);
     }
 
     std::string basename(const std::string & path) {
@@ -653,22 +700,18 @@ class LlamaData {
             return ret;
         }
 
-        const std::string              bn      = basename(model_);
-        const std::vector<std::string> headers = { "--header",
-                                                   "Accept: application/vnd.docker.distribution.manifest.v2+json" };
+        const std::string bn = basename(model_);
         if (string_starts_with(model_, "hf://") || string_starts_with(model_, "huggingface://")) {
             rm_until_substring(model_, "://");
-            ret = huggingface_dl(model_, headers, bn);
+            ret = huggingface_dl(model_, bn);
         } else if (string_starts_with(model_, "hf.co/")) {
             rm_until_substring(model_, "hf.co/");
-            ret = huggingface_dl(model_, headers, bn);
-        } else if (string_starts_with(model_, "ollama://")) {
-            rm_until_substring(model_, "://");
-            ret = ollama_dl(model_, headers, bn);
+            ret = huggingface_dl(model_, bn);
         } else if (string_starts_with(model_, "https://")) {
-            ret = download(model_, headers, bn, true);
-        } else {
-            ret = ollama_dl(model_, headers, bn);
+            ret = download(model_, bn, true);
+        } else {  // ollama:// or nothing
+            rm_until_substring(model_, "://");
+            ret = ollama_dl(model_, bn);
         }
 
         model_ = bn;

ggml/src/ggml-cuda/common.cuh

@@ -46,20 +46,20 @@
 #define GGML_CUDA_CC_VOLTA      700
 #define GGML_CUDA_CC_TURING     750
 #define GGML_CUDA_CC_AMPERE     800
-#define GGML_CUDA_CC_OFFSET_AMD 1000000
+#define GGML_CUDA_CC_OFFSET_AMD 0x1000000
 
 // GCN/CNDA, wave size is 64
-#define GGML_CUDA_CC_GCN4       (GGML_CUDA_CC_OFFSET_AMD + 803)  // Tonga, Fiji, Polaris, minimum for fast fp16
-#define GGML_CUDA_CC_VEGA       (GGML_CUDA_CC_OFFSET_AMD + 900)  // Vega56/64, minimum for fp16 dual issue
-#define GGML_CUDA_CC_VEGA20     (GGML_CUDA_CC_OFFSET_AMD + 906)  // MI50/Radeon VII, minimum for dp4a
-#define GGML_CUDA_CC_CDNA       (GGML_CUDA_CC_OFFSET_AMD + 908)  // MI100, minimum for MFMA, acc registers
-#define GGML_CUDA_CC_CDNA2      (GGML_CUDA_CC_OFFSET_AMD + 910)  // MI210, minimum acc register renameing
-#define GGML_CUDA_CC_CDNA3      (GGML_CUDA_CC_OFFSET_AMD + 942)  // MI300
+#define GGML_CUDA_CC_GCN4       (GGML_CUDA_CC_OFFSET_AMD + 0x803)  // Tonga, Fiji, Polaris, minimum for fast fp16
+#define GGML_CUDA_CC_VEGA       (GGML_CUDA_CC_OFFSET_AMD + 0x900)  // Vega56/64, minimum for fp16 dual issue
+#define GGML_CUDA_CC_VEGA20     (GGML_CUDA_CC_OFFSET_AMD + 0x906)  // MI50/Radeon VII, minimum for dp4a
+#define GGML_CUDA_CC_CDNA       (GGML_CUDA_CC_OFFSET_AMD + 0x908)  // MI100, minimum for MFMA, acc registers
+#define GGML_CUDA_CC_CDNA2      (GGML_CUDA_CC_OFFSET_AMD + 0x910)  // MI210, minimum acc register renameing
+#define GGML_CUDA_CC_CDNA3      (GGML_CUDA_CC_OFFSET_AMD + 0x942)  // MI300
 
 // RNDA removes MFMA, dp4a, xnack, acc registers, wave size is 32
-#define GGML_CUDA_CC_RDNA1      (GGML_CUDA_CC_OFFSET_AMD + 1010) // RX 5000
-#define GGML_CUDA_CC_RDNA2      (GGML_CUDA_CC_OFFSET_AMD + 1030) // RX 6000, minimum for dp4a
-#define GGML_CUDA_CC_RDNA3      (GGML_CUDA_CC_OFFSET_AMD + 1100) // RX 7000, minimum for WMMA
+#define GGML_CUDA_CC_RDNA1      (GGML_CUDA_CC_OFFSET_AMD + 0x1010) // RX 5000
+#define GGML_CUDA_CC_RDNA2      (GGML_CUDA_CC_OFFSET_AMD + 0x1030) // RX 6000, minimum for dp4a
+#define GGML_CUDA_CC_RDNA3      (GGML_CUDA_CC_OFFSET_AMD + 0x1100) // RX 7000, minimum for WMMA
 
 #define GGML_CUDA_CC_QY1        210
 #define GGML_CUDA_CC_QY2        220

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

@@ -119,6 +119,55 @@ static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device)
 #endif
 }
 
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+static int ggml_cuda_parse_id(char devName[]) {
+    // A list of possible Target IDs can be found under the rocclr/clr repo in device.cpp
+    // these values are not stable so this is susceptible to breakage
+    // https://github.com/ROCm/clr/blob/amd-staging/rocclr/device/device.cpp
+    int archMajor = 0x0;
+    int archMinor = 0x0;
+    int archNum = GGML_CUDA_CC_OFFSET_AMD;
+    int archLen = strlen(devName);
+    char archName[archLen + 1];
+
+    // strip leading 'gfx' while copying into our buffer
+    if (archLen > 3) {
+        strcpy(archName, &devName[3]);
+        archLen -= 3;
+    }
+
+    // trim trailing :xnack- or :sramecc- statuses
+    archLen = strcspn(archName, ":");
+    archName[archLen] = '\0';
+
+    // tease out the version information
+    if (archLen > 8) {
+        // versions labeled generic use '-' as delimiter
+        // strip the trailing "-generic" then iterate through what remains
+        if ((strstr(archName, "-generic"))) {
+            archName[archLen - 8] = '\0';
+            char * pch;
+            if ((pch = strtok(archName, "-"))) {
+                archMajor = (int)strtoul(pch, 0, 16);
+                if ((pch = strtok(NULL, "-"))) {
+                    archMinor = 0x10 * (int)strtoul(pch, 0, 16);
+                }
+            }
+        }
+    } else if (archLen >= 3) {
+        // last two digits should be the minor * 0x10 + stepping
+        archMinor = (int)strtoul(&archName[archLen - 2], 0, 16);
+        archName[archLen - 2] = '\0';
+
+        // only the major version remains
+        archMajor = (int)strtoul(archName, 0, 16);
+    }
+    archNum += archMajor * 0x100;
+    archNum += archMinor;
+    return archNum;
+}
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+
 static ggml_cuda_device_info ggml_cuda_init() {
 #ifdef __HIP_PLATFORM_AMD__
     // Workaround for a rocBLAS bug when using multiple graphics cards:
@@ -169,7 +218,6 @@ static ggml_cuda_device_info ggml_cuda_init() {
 
         cudaDeviceProp prop;
         CUDA_CHECK(cudaGetDeviceProperties(&prop, id));
-        GGML_LOG_INFO("  Device %d: %s, compute capability %d.%d, VMM: %s\n", id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no");
 
         info.default_tensor_split[id] = total_vram;
         total_vram += prop.totalGlobalMem;
@@ -178,10 +226,25 @@ static ggml_cuda_device_info ggml_cuda_init() {
         info.devices[id].smpb  = prop.sharedMemPerBlock;
 #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
         info.devices[id].smpbo = prop.sharedMemPerBlock;
-        info.devices[id].cc = 100*prop.major + 10*prop.minor + GGML_CUDA_CC_OFFSET_AMD;
+
+        info.devices[id].cc = ggml_cuda_parse_id(prop.gcnArchName);
+        if ((info.devices[id].cc & 0xff00) == 0x0) {
+            GGML_LOG_WARN("invalid architecture ID received for device %d %s: %s  cc %d.%d\n",
+                            id, prop.name, prop.gcnArchName, prop.major, prop.minor);
+
+            // Fallback to prop.major and prop.minor
+            if (prop.major > 0) {
+                info.devices[id].cc = GGML_CUDA_CC_OFFSET_AMD + prop.major * 0x100;
+                info.devices[id].cc += prop.minor * 0x10;
+            }
+        }
+        GGML_LOG_INFO("  Device %d: %s, %s (0x%x), VMM: %s\n",
+                        id, prop.name, prop.gcnArchName, info.devices[id].cc & 0xffff, device_vmm ? "yes" : "no");
 #else
         info.devices[id].smpbo = prop.sharedMemPerBlockOptin;
         info.devices[id].cc = 100*prop.major + 10*prop.minor;
+        GGML_LOG_INFO("  Device %d: %s, compute capability %d.%d, VMM: %s\n",
+                        id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no");
 #endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
     }
 

src/llama-model-loader.cpp

@@ -819,7 +819,7 @@ void llama_model_loader::init_mappings(bool prefetch, llama_mlocks * mlock_mmaps
         for (const auto & file : files) {
             auto * reg = ggml_backend_dev_backend_reg(ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU));
             auto * is_numa_fn = (decltype(ggml_is_numa) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_is_numa");
-            std::unique_ptr<llama_mmap> mapping(new llama_mmap(file.get(), prefetch ? -1 : 0, is_numa_fn()));
+            std::unique_ptr<llama_mmap> mapping = std::make_unique<llama_mmap>(file.get(), prefetch ? -1 : 0, is_numa_fn());
             mmaps_used.emplace_back(mapping->size(), 0);
             if (mlock_mmaps) {
                 std::unique_ptr<llama_mlock> mlock_mmap(new llama_mlock());

src/llama-vocab.cpp

@@ -1245,8 +1245,13 @@ struct llama_vocab::impl {
 
     std::vector<llama_token> cache_special_tokens;
     std::vector<std::string> cache_token_to_piece; // llama_token_to_piece(special = true);
-
-    std::map<std::pair<std::string, std::string>, int> bpe_ranks;
+    struct pair_hash {
+        size_t operator()(const std::pair<std::string, std::string> & p) const {
+            return std::hash<std::string>{}(p.first) ^  //create some hash for pair
+                   (std::hash<std::string>{}(p.second) << 1);
+        }
+    };
+    std::unordered_map<std::pair<std::string, std::string>, int, pair_hash> bpe_ranks;
 
     // set of all tokens that cause "end of generation"
     std::set<llama_token> special_eog_ids;