Upgrade Space to Qwen3 Coder 30B FP8

README.md

@@ -10,9 +10,11 @@ app_file: app.py
 pinned: false
 ---
 
-This Space serves `Qwen/Qwen3-14B` through a compact OpenAI Chat Completions
-surface. It accepts standard function-tool schemas, disables reasoning traces
-for tool turns, and returns OpenAI-compatible structured tool calls for
+This Space serves `Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8` through a compact
+OpenAI Chat Completions surface. The official 29 GiB FP8 checkpoint fits the
+default 48 GB ZeroGPU slice, activates 3.3B of its 30.5B parameters per token,
+and is specialized for agentic coding. The endpoint accepts standard function-
+tool schemas and returns OpenAI-compatible structured tool calls for
 OpenClaude.
 
 Set `MODEL_ID` only when deliberately testing another compatible model; the

app.py

@@ -31,10 +31,14 @@ from transformers import (
 from tool_calls import extract_tool_call, has_complete_tool_call
 
 
-# Qwen3-14B fits comfortably in the Space's ZeroGPU allocation and, unlike the
-# previous coding model, has a dedicated chat template for OpenAI-style tool
-# calls. Keep MODEL_ID as an override so the Space remains easy to roll back.
-MODEL = os.getenv("MODEL_ID", "Qwen/Qwen3-14B")
+# Qwen3-Coder-30B-A3B is the strongest coding-specialized Qwen checkpoint that
+# fits the default 48 GB ZeroGPU slice and the Space's 50 GB disk.  The official
+# FP8 checkpoint is about 29 GiB, activates only 3.3B parameters per token, and
+# has a dedicated chat template for OpenAI-style tool calls.  Keep MODEL_ID as
+# an override so the Space remains easy to roll back.
+MODEL = os.getenv(
+    "MODEL_ID", "Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8"
+)
 MAX_CONTEXT_TOKENS = int(os.getenv("MAX_CONTEXT_TOKENS", "24000"))
 MAX_NEW_TOKENS = int(os.getenv("MAX_NEW_TOKENS", "768"))
 TOOL_PROTOCOL_MARKER = "OPENAI TOOL CALL FORMAT — MANDATORY"
@@ -47,7 +51,9 @@ tokenizer.truncation_side = "left"
 print(f"Loading {MODEL} on ZeroGPU during startup...", flush=True)
 model = AutoModelForCausalLM.from_pretrained(
     MODEL,
-    dtype=torch.bfloat16,
+    # Preserve the checkpoint's fine-grained FP8 weights.  Forcing BF16 here
+    # would roughly double memory use and no longer fit the default GPU slice.
+    torch_dtype="auto",
     low_cpu_mem_usage=True,
 )
 model = model.to("cuda")
@@ -303,10 +309,13 @@ def _completion_payload(request: ChatCompletionRequest) -> dict[str, Any]:
     if request.model not in {
         MODEL,
         "qwen-coder",
+        "qwen3-coder",
+        "qwen3-coder-30b-a3b",
         "qwen3",
         "qwen3-14b",
-        # Accept the previous backend identifier during the local proxy's
-        # rolling restart so an in-flight OpenClaude session is not rejected.
+        # Accept previous backend identifiers during the local proxy's rolling
+        # restart so an in-flight OpenClaude session is not rejected.
+        "Qwen/Qwen3-14B",
         "Qwen/Qwen2.5-Coder-14B-Instruct",
         "qwen2.5-coder-14b",
     }:
@@ -391,7 +400,7 @@ demo = gr.Interface(
         gr.Number(value=512, label="Max Tokens"),
     ],
     outputs="text",
-    title="Qwen Coder ZeroGPU Backend",
+    title="Qwen3 Coder 30B ZeroGPU Backend",
 )
 
 class OpenAIRouteMiddleware(BaseHTTPMiddleware):

requirements.txt

@@ -1,5 +1,5 @@
 fastapi
 pydantic
-transformers
-torch
-accelerate
+transformers>=4.52.3
+torch>=2.8.0
+accelerate>=1.2.0