Upload app.py with huggingface_hub

app.py

@@ -0,0 +1,101 @@
+import spaces  # noqa: E402  (must be imported before torch/transformers)
+# v2: added kernels dependency for FP8 inference
+
+import json
+import re
+import time
+
+import gradio as gr
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+MODEL_ID = "Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8"
+
+tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
+model = AutoModelForCausalLM.from_pretrained(
+    MODEL_ID,
+    dtype="auto",
+).to("cuda")
+model.eval()
+
+TOOL_CALL_RE = re.compile(r"<tool_call>\s*<function=(?P<name>[^>]+)>(?P<body>.*?)</function>\s*</tool_call>", re.DOTALL)
+PARAM_RE = re.compile(r"<parameter=(?P<name>[^>]+)>(?P<value>.*?)</parameter>", re.DOTALL)
+
+
+def _parse_tool_calls(text: str):
+    """Extract <tool_call> blocks from generated text into OpenAI-style tool_calls."""
+    tool_calls = []
+    for match in TOOL_CALL_RE.finditer(text):
+        name = match.group("name").strip()
+        body = match.group("body")
+        arguments = {}
+        for pmatch in PARAM_RE.finditer(body):
+            pname = pmatch.group("name").strip()
+            pvalue = pmatch.group("value").strip("\n")
+            arguments[pname] = pvalue
+        tool_calls.append(
+            {
+                "id": f"call_{len(tool_calls)}_{int(time.time() * 1000)}",
+                "type": "function",
+                "function": {"name": name, "arguments": json.dumps(arguments)},
+            }
+        )
+    leftover = TOOL_CALL_RE.sub("", text).strip()
+    return leftover, tool_calls
+
+
+@spaces.GPU(duration=45)
+def chat_completions(messages_json: str, tools_json: str = "", max_new_tokens: int = 2048) -> str:
+    """Run one chat-completion turn against Qwen3-Coder-30B-A3B-Instruct-FP8.
+
+    messages_json: JSON-encoded list of {"role", "content"} (and optionally
+        "tool_calls" / "tool_call_id" for prior turns), OpenAI chat format.
+    tools_json: JSON-encoded list of OpenAI-style tool/function definitions,
+        or an empty string for no tools.
+    max_new_tokens: generation budget for this turn.
+
+    Returns a JSON-encoded object: {"content": str, "tool_calls": [...]}.
+    """
+    messages = json.loads(messages_json)
+    tools = json.loads(tools_json) if tools_json else None
+
+    prompt = tokenizer.apply_chat_template(
+        messages,
+        tools=tools,
+        add_generation_prompt=True,
+        tokenize=False,
+    )
+    inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
+
+    with torch.no_grad():
+        output_ids = model.generate(
+            **inputs,
+            max_new_tokens=max_new_tokens,
+            do_sample=False,
+        )
+
+    generated = output_ids[0][inputs["input_ids"].shape[1]:]
+    text = tokenizer.decode(generated, skip_special_tokens=True)
+
+    content, tool_calls = _parse_tool_calls(text)
+    return json.dumps({"content": content, "tool_calls": tool_calls})
+
+
+demo = gr.Interface(
+    fn=chat_completions,
+    inputs=[
+        gr.Text(label="messages_json"),
+        gr.Text(label="tools_json", value=""),
+        gr.Number(label="max_new_tokens", value=2048, precision=0),
+    ],
+    outputs=gr.Text(label="completion_json"),
+    title="Qwen3-Coder-30B-A3B-Instruct-FP8 — chat completions",
+    description=(
+        "Internal inference endpoint for the Coding Model project. "
+        "Called via the local OpenAI-compatible shim, not directly by end users."
+    ),
+    api_name="chat_completions",
+)
+
+if __name__ == "__main__":
+    demo.launch(mcp_server=True)