Both model sizes on ZeroGPU

README.md

@@ -14,6 +14,7 @@ models:
   - TigreGotico/jevasio-0.8b
   - TigreGotico/jevasio-4b
   - Qwen/Qwen3.5-0.8B-Base
+  - Qwen/Qwen3.5-4B-Base
 datasets:
   - TigreGotico/jevasio-decisions
 tags:
@@ -28,8 +29,9 @@ Jevásio is a decision model for European Portuguese (pt-PT). It reads a state,
 set of typed questions about it. It returns a probability for each option of each question in one forward pass. It
 does not generate text.
 
-This Space runs [Jevásio-0.8B](https://huggingface.co/TigreGotico/jevasio-0.8b) on a CPU, in fp32, so one request
-takes a few seconds. [Jevásio-4B](https://huggingface.co/TigreGotico/jevasio-4b) is more accurate and needs a GPU.
+This Space runs both sizes on ZeroGPU, in bf16: [Jevásio-4B](https://huggingface.co/TigreGotico/jevasio-4b), the more
+accurate one, and [Jevásio-0.8B](https://huggingface.co/TigreGotico/jevasio-0.8b), the smaller one. Pick one with the
+model selector.
 Both models are [Kev](https://github.com/jaredpalmer/kev) fine-tuned on the
 [TigreGotico/jevasio-decisions](https://huggingface.co/datasets/TigreGotico/jevasio-decisions) dataset.
 

app.py

@@ -1,9 +1,9 @@
-"""Jevásio on Hugging Face Spaces (Gradio, CPU).
+"""Jevásio on Hugging Face Spaces (Gradio, ZeroGPU).
 
 A state and a set of typed questions in, a probability per option out, in one forward pass. The model is loaded through
 Kev's own loader (kev.checkpoint), the same path `python -m kev.serve --run TigreGotico/jevasio-0.8b` takes, and
-requests go through kev.api and kev.model.admit exactly as /v1/systemone does. On CPU the weights stay in fp32
-(KEV_DTYPE=bf16 overrides it).
+requests go through kev.api and kev.model.admit exactly as /v1/systemone does. Both sizes load in bf16 at start-up, and
+each forward pass runs inside @spaces.GPU, so a request holds a GPU only for that pass.
 """
 import html
 import json
@@ -12,6 +12,7 @@ import time
 from dataclasses import replace
 
 import gradio as gr
+import spaces
 import torch
 from pydantic import ValidationError
 
@@ -19,22 +20,28 @@ from kev.api import SystemOneRequest, output_tokens, render, to_answers, to_reco
 from kev.checkpoint import Checkpoint, LoadOptions
 from kev.model import admit
 
-MODEL_REPO = os.environ.get("JEVASIO_MODEL", "TigreGotico/jevasio-0.8b")
+MODELS = {"Jevásio-4B": "TigreGotico/jevasio-4b", "Jevásio-0.8B": "TigreGotico/jevasio-0.8b"}
+DEFAULT = "Jevásio-4B"
 LINK_08B = "https://huggingface.co/TigreGotico/jevasio-0.8b"
 LINK_4B = "https://huggingface.co/TigreGotico/jevasio-4b"
 LINK_DATA = "https://huggingface.co/datasets/TigreGotico/jevasio-decisions"
 
-torch.set_num_threads(os.cpu_count() or 1)
+LOADED = {}
+for name, repo in MODELS.items():
+    t0 = time.perf_counter()
+    checkpoint = Checkpoint(repo)
+    opts = LoadOptions.from_env()
+    if opts.dtype is None:
+        opts = replace(opts, dtype=torch.bfloat16)
+    LOADED[name] = checkpoint.load("cuda", opts)
+    tok, model = LOADED[name]
+    print(f"[jevasio] {repo}: base={checkpoint.meta.base}@{checkpoint.meta.base_revision} dtype={model.dtype} "
+          f"temperature={model.head.temperature:.2f} loaded in {time.perf_counter() - t0:.0f} s", flush=True)
 
-t0 = time.perf_counter()
-CHECKPOINT = Checkpoint(MODEL_REPO)
-opts = LoadOptions.from_env()
-if opts.dtype is None:
-    opts = replace(opts, dtype=torch.float32)
-TOKENIZER, MODEL = CHECKPOINT.load("cpu", opts)
-LOAD_SECONDS = time.perf_counter() - t0
-print(f"[jevasio] {MODEL_REPO}: base={CHECKPOINT.meta.base}@{CHECKPOINT.meta.base_revision} dtype={MODEL.dtype} "
-      f"temperature={MODEL.head.temperature:.2f} loaded in {LOAD_SECONDS:.0f} s", flush=True)
+
+@spaces.GPU(duration=30)
+def forward(name, enc):
+    return [p.float().cpu() for p in LOADED[name][1].probs(enc)]
 
 
 class RequestError(ValueError):
@@ -50,19 +57,22 @@ def parse_request(payload):
         raise RequestError(f"Pedido inválido em `{where}`: {first.get('msg')}") from None
 
 
-def systemone(req):
+def systemone(req, name=DEFAULT):
     """The /v1/systemone response body kev.serve returns for this request, plus the latency of the forward pass."""
+    if name not in LOADED:
+        raise RequestError(f"Modelo desconhecido: {name}. Opções: {', '.join(MODELS)}.")
+    tokenizer, model = LOADED[name]
     rec, meta = to_record(req)
     try:
-        enc = admit(MODEL, TOKENIZER, rec)
+        enc = admit(model, tokenizer, rec)
     except ValueError as e:
         raise RequestError(str(e)) from None
     start = time.perf_counter()
-    probs = MODEL.probs(enc)
+    probs = forward(name, enc)
     ms = (time.perf_counter() - start) * 1000
     answers = to_answers([p.tolist() for p in probs], meta)
-    return {"model": MODEL_REPO, "answers": answers,
-            "usage": {"input_tokens": len(enc["ids"]), "output_tokens": output_tokens(TOKENIZER, answers)},
+    return {"model": MODELS[name], "answers": answers,
+            "usage": {"input_tokens": len(enc["ids"]), "output_tokens": output_tokens(tokenizer, answers)},
             "latency_ms": round(ms, 1)}
 
 
@@ -91,7 +101,7 @@ def render_answers(req, resp):
     usage = resp["usage"]
     parts = ['<div style="font-size:12px;color:var(--body-text-color-subdued);margin:6px 0 10px">'
              f'{len(resp["answers"])} pergunta(s) · {usage["input_tokens"]} tokens de entrada · '
-             f'{resp["latency_ms"]:.0f} ms em CPU · nenhum texto gerado</div>']
+             f'{resp["model"].split("/")[1]} · {resp["latency_ms"]:.0f} ms · nenhum texto gerado</div>']
     for qid, a in resp["answers"].items():
         instructions = render(req.questions[qid].instructions)
         if a["type"] == "noul":
@@ -167,18 +177,18 @@ def build_request(state, choice_id, choice_instr, choice_opts, noul_id, noul_ins
     return {"state": parse_state(state), "model": "kev-latest", "questions": questions}
 
 
-def decide(state, choice_id, choice_instr, choice_opts, noul_id, noul_instr, score_id, score_instr, score_levels):
+def decide(name, state, choice_id, choice_instr, choice_opts, noul_id, noul_instr, score_id, score_instr, score_levels):
     """Answer the builder's questions about the state. Returns the rendered answers, the request and the response."""
     try:
         payload = build_request(state, choice_id, choice_instr, choice_opts, noul_id, noul_instr, score_id, score_instr, score_levels)
         req = parse_request(payload)
-        resp = systemone(req)
+        resp = systemone(req, name)
     except RequestError as e:
         return error_html(str(e)), {}, {"error": str(e)}
     return render_answers(req, resp), payload, resp
 
 
-def decide_raw(request_json):
+def decide_raw(name, request_json):
     """Answer a raw System One request (JSON text). Returns a status line and the response or an error object."""
     try:
         payload = json.loads(request_json or "")
@@ -189,7 +199,7 @@ def decide_raw(request_json):
         message = 'O pedido tem de ser um objeto JSON com "state" e "questions".'
         return f"**Erro:** {message}", {"error": message}
     try:
-        resp = systemone(parse_request(payload))
+        resp = systemone(parse_request(payload), name)
     except RequestError as e:
         return f"**Erro:** {e}", {"error": str(e)}
     return f"Respondido em {resp['latency_ms']:.0f} ms.", resp
@@ -263,8 +273,7 @@ INTRO = f"""# Jevásio
 Modelo de decisão para português europeu. Recebe um **estado** (um texto ou um objeto JSON) e perguntas tipadas, e
 devolve uma probabilidade para cada opção de cada pergunta, numa só passagem pelo modelo. Não gera texto.
 
-Este Space corre o [Jevásio-0.8B]({LINK_08B}) em CPU, por isso cada pedido demora alguns segundos. Existe também o
-[Jevásio-4B]({LINK_4B}), mais exato, que precisa de GPU.
+Escolha o modelo: o [Jevásio-4B]({LINK_4B}) é mais exato, o [Jevásio-0.8B]({LINK_08B}) é mais pequeno e rápido.
 """
 
 ABOUT = f"""Há três tipos de pergunta: `choice` (escolher uma opção com nome), `noul` (sim ou não) e `score` (um nível numa
@@ -273,7 +282,7 @@ valores em português, como nos dados de treino.
 
 O Jevásio é o [Kev](https://github.com/jaredpalmer/kev) afinado com o conjunto
 [TigreGotico/jevasio-decisions]({LINK_DATA}). A probabilidade usa a temperatura de calibração guardada no modelo
-(T = {MODEL.head.temperature:.2f}). Antes de definir um limiar, meça o modelo com os seus próprios dados.
+(T = {LOADED["Jevásio-4B"][1].head.temperature:.2f} no 4B e {LOADED["Jevásio-0.8B"][1].head.temperature:.2f} no 0.8B). Antes de definir um limiar, meça o modelo com os seus próprios dados.
 
 Licença CC BY-NC-SA 4.0: só para investigação não comercial. Não use o modelo em decisões com consequências legais,
 médicas, financeiras ou laborais sem revisão humana.
@@ -293,6 +302,7 @@ first = EXAMPLES[0]
 with gr.Blocks(title="Jevásio") as demo:
     with gr.Column(elem_id="col-container"):
         gr.Markdown(INTRO)
+        model_choice = gr.Radio(list(MODELS), value=DEFAULT, label="Modelo")
         with gr.Tabs():
             with gr.Tab("Perguntas"):
                 with gr.Row():
@@ -330,7 +340,7 @@ with gr.Blocks(title="Jevásio") as demo:
                 inputs = [state_box, choice_id, choice_instr, choice_opts, noul_id, noul_instr, score_id, score_instr, score_levels]
                 gr.Examples(examples=[example_row(ex) for ex in EXAMPLES], example_labels=[ex["name"] for ex in EXAMPLES],
                             inputs=inputs, label="Exemplos")
-                run_btn.click(fn=decide, inputs=inputs, outputs=[answers_html, request_json, response_json], api_name="decide")
+                run_btn.click(fn=decide, inputs=[model_choice, *inputs], outputs=[answers_html, request_json, response_json], api_name="decide")
 
             with gr.Tab("Avançado"):
                 gr.Markdown("Cole um pedido em JSON no formato System One (`POST /v1/systemone`) e veja a resposta em JSON.")
@@ -341,7 +351,7 @@ with gr.Blocks(title="Jevásio") as demo:
                     with gr.Column(scale=1):
                         raw_status = gr.Markdown()
                         raw_out = gr.JSON(label="Resposta (JSON)")
-                raw_btn.click(fn=decide_raw, inputs=[raw_in], outputs=[raw_status, raw_out], api_name="systemone")
+                raw_btn.click(fn=decide_raw, inputs=[model_choice, raw_in], outputs=[raw_status, raw_out], api_name="systemone")
 
             with gr.Tab("Sobre"):
                 gr.Markdown(ABOUT)

requirements.txt

@@ -1,3 +1,3 @@
---extra-index-url https://download.pytorch.org/whl/cpu
-torch==2.8.0+cpu
+torch==2.8.0
+spaces
 kev @ git+https://github.com/jaredpalmer/kev@1f04cfa28e6a708485b28a9585010b9e436b776e