(Studio2) Refactors SD pipeline to rely on turbine-models pipeline, fixes to LLM, gitignore

.github/workflows/test-studio.yml

@@ -81,6 +81,4 @@ jobs:
         source shark.venv/bin/activate
         pip install -r requirements.txt --no-cache-dir
         pip install -e .
-        pip uninstall -y torch
-        pip install torch==2.1.0+cpu -f https://download.pytorch.org/whl/torch_stable.html
         python apps/shark_studio/tests/api_test.py

.gitignore

@@ -164,14 +164,15 @@ cython_debug/
 # vscode related
 .vscode
 
-# Shark related artefacts
+# Shark related artifacts
 *venv/
 shark_tmp/
 *.vmfb
 .use-iree
 tank/dict_configs.py
 *.csv
 reproducers/
+apps/shark_studio/web/configs
 
 # ORT related artefacts
 cache_models/
@@ -188,6 +189,11 @@ variants.json
 # models folder
 apps/stable_diffusion/web/models/
 
+# model artifacts (SHARK)
+*.tempfile
+*.mlir
+*.vmfb
+
 # Stencil annotators.
 stencil_annotator/
 

apps/shark_studio/api/llm.py

@@ -155,7 +155,9 @@ def __init__(
                 use_auth_token=hf_auth_token,
             )
         elif not os.path.exists(self.tempfile_name):
-            self.torch_ir, self.tokenizer = llm_model_map[model_name]["initializer"](
+            self.torch_ir, self.tokenizer = llm_model_map[self.hf_model_name][
+                "initializer"
+            ](
                 self.hf_model_name,
                 hf_auth_token,
                 compile_to="torch",
@@ -258,8 +260,7 @@ def format_out(results):
 
             history.append(format_out(token))
             while (
-                format_out(token)
-                != llm_model_map["meta-llama/Llama-2-7b-chat-hf"]["stop_token"]
+                format_out(token) != llm_model_map[self.hf_model_name]["stop_token"]
                 and len(history) < self.max_tokens
             ):
                 dec_time = time.time()
@@ -273,10 +274,7 @@ def format_out(results):
 
             self.prev_token_len = token_len + len(history)
 
-            if (
-                format_out(token)
-                == llm_model_map["meta-llama/Llama-2-7b-chat-hf"]["stop_token"]
-            ):
+            if format_out(token) == llm_model_map[self.hf_model_name]["stop_token"]:
                 break
 
         for i in range(len(history)):
@@ -310,7 +308,7 @@ def chat_hf(self, prompt):
                 self.first_input = False
 
             history.append(int(token))
-            while token != llm_model_map["meta-llama/Llama-2-7b-chat-hf"]["stop_token"]:
+            while token != llm_model_map[self.hf_model_name]["stop_token"]:
                 dec_time = time.time()
                 result = self.hf_mod(token.reshape([1, 1]), past_key_values=pkv)
                 history.append(int(token))
@@ -321,7 +319,7 @@ def chat_hf(self, prompt):
 
             self.prev_token_len = token_len + len(history)
 
-            if token == llm_model_map["meta-llama/Llama-2-7b-chat-hf"]["stop_token"]:
+            if token == llm_model_map[self.hf_model_name]["stop_token"]:
                 break
         for i in range(len(history)):
             if type(history[i]) != int: