Add support for http, https and file pulls

Summary by Sourcery Add support for pulling models using http, https, and file URLs, enabling direct execution of models from web or local sources. Update documentation and add tests to cover the new functionality. New Features: Add support for pulling models using http, https, and file URLs, allowing models to be run directly from web sources or local files. Enhancements: Refactor the symlink creation process to use os.symlink directly instead of run_cmd. Documentation: Update documentation to include new URL syntax support for http, https, and file protocols, explaining how models can be pulled from these sources. Tests: Add system tests to verify the functionality of pulling models using file and https URLs, ensuring they can be listed and removed correctly. Signed-off-by: Daniel J Walsh <[email protected]>
containers · Nov 18, 2024 · 45911be · 45911be
1 parent fb5ee93
commit 45911be
Show file tree

Hide file tree

Showing 6 changed files with 94 additions and 30 deletions.
diff --git a/docs/ramalama.1.md b/docs/ramalama.1.md
@@ -40,13 +40,21 @@ RamaLama supports multiple AI model registries types called transports. Supporte
 | OCI Container Registries | [`opencontainers.org`](https://opencontainers.org)|
 ||Examples: [`quay.io`](https://quay.io),  [`Docker Hub`](https://docker.io), and [`Artifactory`](https://artifactory.com)|
 
+RamaLama can also pull directly using URL syntax.
+
+http://, https:// and file://.
+
+This means if a model is on a web site or even on your local system, you can run it directly.
+
 RamaLama uses the Ollama registry transport by default. The default can be overridden in the ramalama.conf file or use the RAMALAMA_TRANSPORTS
 environment. `export RAMALAMA_TRANSPORT=huggingface` Changes RamaLama to use huggingface transport.
 
-Individual model transports can be modifies when specifying a model via the `huggingface://`, `oci://`, or `ollama://` prefix.
+Individual model transports can be modifies when specifying a model via the `huggingface://`, `oci://`, `ollama://`, `https://`, `http://`, `file://` prefix.
 
 ramalama pull `huggingface://`afrideva/Tiny-Vicuna-1B-GGUF/tiny-vicuna-1b.q2_k.gguf
 
+ramalama run `file://`$HOME/granite-7b-lab-Q4_K_M.gguf
+
 To make it easier for users, RamaLama uses shortname files, which container
 alias names for fully specified AI Models allowing users to specify the shorter
 names when referring to models. RamaLama reads shortnames.conf files if they

diff --git a/ramalama/cli.py b/ramalama/cli.py
@@ -17,6 +17,7 @@
 )
 from ramalama.oci import OCI
 from ramalama.ollama import Ollama
+from ramalama.url import URL
 from ramalama.shortnames import Shortnames
 from ramalama.toml_parser import TOMLParser
 from ramalama.version import version, print_version
@@ -430,7 +431,10 @@ def _list_models(args):
     # Collect model data
     for path in list_files_by_modification():
         if path.is_symlink():
-            name = str(path).replace("/", "://", 1)
+            if str(path).startswith("file/"):
+                name = str(path).replace("/", ":///", 1)
+            else:
+                name = str(path).replace("/", "://", 1)
             file_epoch = path.lstat().st_mtime
             modified = int(time.time() - file_epoch)
             size = get_size(path)
@@ -762,6 +766,8 @@ def New(model, args):
         return Ollama(model)
     if model.startswith("oci://") or model.startswith("docker://"):
         return OCI(model, args.engine)
+    if model.startswith("http://") or model.startswith("https://") or model.startswith("file://"):
+        return URL(model)
 
     transport = config.get("transport", "ollama")
     if transport == "huggingface":

diff --git a/ramalama/huggingface.py b/ramalama/huggingface.py
@@ -37,7 +37,7 @@ def __init__(self, model):
         model = model.removeprefix("huggingface://")
         model = model.removeprefix("hf://")
         super().__init__(model)
-        self.type = "HuggingFace"
+        self.type = "huggingface"
         split = self.model.rsplit("/", 1)
         self.directory = split[0] if len(split) > 1 else ""
         self.filename = split[1] if len(split) > 1 else split[0]
@@ -61,16 +61,6 @@ def logout(self, args):
             conman_args.extend(["--token", args.token])
         self.exec(conman_args)
 
-    def path(self, args):
-        return self.model_path(args)
-
-    def exists(self, args):
-        model_path = self.model_path(args)
-        if not os.path.exists(model_path):
-            return None
-
-        return model_path
-
     def pull(self, args):
         model_path = self.model_path(args)
         directory_path = os.path.join(args.store, "repos", "huggingface", self.directory, self.filename)
@@ -93,13 +83,12 @@ def pull(self, args):
         if os.path.exists(target_path) and verify_checksum(target_path):
             relative_target_path = os.path.relpath(target_path, start=os.path.dirname(model_path))
             if not self.check_valid_model_path(relative_target_path, model_path):
-                run_cmd(["ln", "-sf", relative_target_path, model_path], debug=args.debug)
+                os.symlink(relative_target_path, model_path)
             return model_path
 
         # Download the model file to the target path
         url = f"https://huggingface.co/{self.directory}/resolve/main/{self.filename}"
         download_file(url, target_path, headers={}, show_progress=True)
-
         if not verify_checksum(target_path):
             print(f"Checksum mismatch for {target_path}, retrying download...")
             os.remove(target_path)
@@ -112,8 +101,7 @@ def pull(self, args):
             # Symlink is already correct, no need to update it
             return model_path
 
-        run_cmd(["ln", "-sf", relative_target_path, model_path], debug=args.debug)
-
+        os.symlink(relative_target_path, model_path)
         return model_path
 
     def push(self, source, args):
@@ -137,12 +125,6 @@ def push(self, source, args):
         )
         return proc.stdout.decode("utf-8")
 
-    def model_path(self, args):
-        return os.path.join(args.store, "models", "huggingface", self.directory, self.filename)
-
-    def check_valid_model_path(self, relative_target_path, model_path):
-        return os.path.exists(model_path) and os.readlink(model_path) == relative_target_path
-
     def exec(self, args):
         try:
             exec_cmd(args, args.debug)

diff --git a/ramalama/model.py b/ramalama/model.py
@@ -48,9 +48,6 @@ def login(self, args):
     def logout(self, args):
         raise NotImplementedError(f"ramalama logout for {self.type} not implemented")
 
-    def path(self, source, args):
-        raise NotImplementedError(f"ramalama path for {self.type} not implemented")
-
     def pull(self, args):
         raise NotImplementedError(f"ramalama pull for {self.type} not implemented")
 
@@ -67,7 +64,7 @@ def is_symlink_to(self, file_path, target_path):
         return False
 
     def garbage_collection(self, args):
-        repo_paths = ["huggingface", "oci", "ollama"]
+        repo_paths = ["huggingface", "oci", "ollama", "file", "http", "https"]
         for repo in repo_paths:
             repo_dir = f"{args.store}/repos/{repo}"
             model_dir = f"{args.store}/models/{repo}"
@@ -102,9 +99,6 @@ def remove(self, args):
 
         self.garbage_collection(args)
 
-    def model_path(self, args):
-        raise NotImplementedError(f"model_path for {self.type} not implemented")
-
     def _image(self, args):
         if args.image != default_image():
             return args.image
@@ -343,6 +337,21 @@ def kube(self, model, args, exec_args):
         kube = Kube(model, args, exec_args)
         kube.generate()
 
+    def path(self, args):
+        return self.model_path(args)
+
+    def model_path(self, args):
+        return os.path.join(args.store, "models", self.type, self.directory, self.filename)
+
+    def exists(self, args):
+        model_path = self.model_path(args)
+        if not os.path.exists(model_path):
+            return None
+
+        return model_path
+
+    def check_valid_model_path(self, relative_target_path, model_path):
+        return os.path.exists(model_path) and os.readlink(model_path) == relative_target_path
 
 def get_gpu():
     i = 0

diff --git a/ramalama/url.py b/ramalama/url.py
@@ -0,0 +1,43 @@
+import os
+from ramalama.common import download_file
+from ramalama.model import Model
+
+
+class URL(Model):
+    def __init__(self, model):
+        self.type = ""
+        for prefix in ["file", "http", "https"]:
+            if model.startswith(f"{prefix}://"):
+                self.type = prefix
+                model = model.removeprefix(f"{prefix}://")
+                break
+
+        super().__init__(model)
+        split = self.model.rsplit("/", 1)
+        self.directory = split[0].removeprefix("/") if len(split) > 1 else ""
+        self.filename = split[1] if len(split) > 1 else split[0]
+
+    def pull(self, args):
+        model_path = self.model_path(args)
+        directory_path = os.path.join(args.store, "repos", self.type, self.directory, self.filename)
+        os.makedirs(directory_path, exist_ok=True)
+
+        symlink_dir = os.path.dirname(model_path)
+        os.makedirs(symlink_dir, exist_ok=True)
+
+        target_path = os.path.join(directory_path, self.filename)
+
+        if self.type == "file":
+            os.symlink(self.model, os.path.join(symlink_dir, self.filename))
+            os.symlink(self.model, target_path)
+        else:
+            url = self.type + "://" + self.model
+            # Download the model file to the target path
+            download_file(url, target_path, headers={}, show_progress=True)
+            relative_target_path = os.path.relpath(target_path, start=os.path.dirname(model_path))
+            if self.check_valid_model_path(relative_target_path, model_path):
+                # Symlink is already correct, no need to update it
+                return model_path
+            os.symlink(relative_target_path, model_path)
+
+        return model_path
diff --git a/test/system/050-pull.bats b/test/system/050-pull.bats
@@ -60,6 +60,22 @@ load setup_suite
     run_ramalama rm oci://quay.io/mmortari/gguf-py-example:v1
 }
 
+@test "ramalama file URL" {
+      model=$RAMALAMA_TMPDIR/mymodel.gguf
+      touch $model
+      file_url=file://${model}
+      https_url=https://github.com/containers/ramalama/blob/main/README.md
+
+      for url in $file_url $https_url; do
+          run_ramalama pull $url
+          run_ramalama list
+          is "$output" ".*$url" "URL exists"
+          run_ramalama rm $url
+          run_ramalama list
+          assert "$output" !~ ".*$url" "URL no longer exists"
+      done
+}
+
 @test "ramalama use registry" {
     skip_if_darwin
     skip_if_docker