"""NeoJev on NVIDIA GPUs (PyTorch + transformers). Same API and method as decider.py (MLX). Weights: dowands/neojev (the official Qwen/Qwen3.8-27B-FP8 checkpoint cut to its first 56 decoder layers; the full Qwen/Qwen3.8-27B-FP8 also works, only its first 56 layers are loaded). Needs a GPU with FP8 support for the FP8 checkpoint (Ada / Hopper or newer) and about 30 GB of memory. Install flash-linear-attention and causal-conv1d for the fast Gated DeltaNet kernels; without them transformers falls back to a much slower pure-PyTorch path. from decider_cuda import Decider d = Decider("dowands/neojev") d.ask("Customer says the refund never arrived.", {"queue": {"type": "choice", "options": ["Billing", "Technical"], "prompt": "Which team?"}}) d.ask("", {"screen": {"type": "choice", "options": ["main menu", "map", "settings"]}}, image=open("screenshot.png", "rb").read()) # needs vision_common.py next to this file """ import copy import json import os from collections import OrderedDict import numpy as np import torch from transformers import AutoConfig, AutoModelForImageTextToText, AutoTokenizer from transformers.cache_utils import DynamicCache LETTERS = "ABCDEFGHIJKLMNOP" VOCAB = list(LETTERS) + ["Yes", "No"] SYSTEM = ("You are a decision engine. You get a list of QUESTIONS and a STATE. For each question, " "answer with exactly one option label.") def _block(i, name, spec): t = spec["type"] if t == "noul": return f"Q{i} ({name}): Is this statement true? \"{spec['statement']}\"\nOPTIONS: Yes / No", ["Yes", "No"], ["Yes", "No"] opts = spec["options"] if t == "choice" else spec["levels"] if len(opts) > len(LETTERS): raise ValueError(f"question {name!r}: at most {len(LETTERS)} options are supported") labels = list(LETTERS[: len(opts)]) lines = "\n".join(f"{l}. {o}" for l, o in zip(labels, opts)) return f"Q{i} ({name}): {spec.get('prompt', 'Which option fits best?')}\nOPTIONS:\n{lines}", labels, opts def _dequant_rows(lin, ids): """Rows of the LM head as float32, whether the checkpoint stores it in bf16 or block-FP8.""" w = lin.weight if w.dtype in (torch.float8_e4m3fn, torch.float8_e5m2): scale = getattr(lin, "weight_scale_inv", None) rows = w[ids].to(torch.float32) if scale is not None: # block-wise scales, typically 128 x 128 bs_r = w.shape[0] // scale.shape[0] bs_c = w.shape[1] // scale.shape[1] s = scale.to(torch.float32)[ids // bs_r] rows = rows * s.repeat_interleave(bs_c, dim=1)[:, : w.shape[1]] return rows return w[ids].to(torch.float32) def _patch_fp8_dense(): """transformers 5.16 FP8 quantizer looks up an MoE experts implementation that dense models don't set.""" try: from transformers.integrations.finegrained_fp8 import FP8Experts FP8Experts._impl_tp_layer_overrides.setdefault(None, {}) except Exception: pass class Decider: def __init__(self, model_path="dowands/neojev", head=None, n_layers=56, token_budget=160, device="cuda", vision_head=None): _patch_fp8_dense() if head is None: # the readout head ships next to the weights (local folder or Hugging Face repo) local = os.path.join(model_path, "decider_head.npz") if os.path.exists(local): head = local else: from huggingface_hub import hf_hub_download head = hf_hub_download(model_path, "decider_head.npz") if vision_head is None: # optional second head for image calls, next to the first one local = os.path.join(model_path, "decider_head_vision.npz") if os.path.exists(local): vision_head = local elif not os.path.isdir(model_path): try: from huggingface_hub import hf_hub_download vision_head = hf_hub_download(model_path, "decider_head_vision.npz") except Exception: vision_head = None cfg = AutoConfig.from_pretrained(model_path) qc = getattr(cfg, "quantization_config", None) if isinstance(qc, dict) and qc.get("modules_to_not_convert"): # the FP8 checkpoint lists MoE router names ("...mlp.gate") that prefix-match the dense "mlp.gate_proj", # which would then load raw FP8 bytes as bf16 and drop their scales; keep gate_proj quantized qc["modules_to_not_convert"] = [m for m in qc["modules_to_not_convert"] if not m.endswith(".mlp.gate")] tc = cfg.text_config tc.num_hidden_layers = n_layers tc.layer_types = tc.layer_types[:n_layers] self.model = AutoModelForImageTextToText.from_pretrained(model_path, config=cfg, dtype="auto", device_map=device) self.model.eval() self.tok = AutoTokenizer.from_pretrained(model_path) self.image_token_id = getattr(cfg, "image_token_id", None) self.text = self.model.model.language_model self.final_norm = self.text.norm self.text.norm = torch.nn.Identity() # we read the pre-norm residual stream of the last kept layer self.device = device self.budget = token_budget t = self.tok.apply_chat_template( [{"role": "system", "content": SYSTEM}, {"role": "user", "content": "<>"}], tokenize=False, add_generation_prompt=True, enable_thinking=False) self.head, self.tail = t.split("<>") # text calls read out through `head`; image calls through `vision_head` when there is one self.text_head = self._load_head(head) self.vision_head = self._load_head(vision_head) if vision_head else self.text_head ids = [] for l in VOCAB: e = self.tok.encode(" " + l, add_special_tokens=False) ids.append(e[0] if len(e) == 1 else self.tok.encode(l, add_special_tokens=False)[0]) self.W = _dequant_rows(self.model.lm_head, torch.tensor(ids, device=device)) # Qwen3.5 RMSNorm in transformers computes x_hat * (1 + weight): the checkpoint stores weight - 1 self.norm_w = 1.0 + self.final_norm.weight.to(torch.float32) self.eps = self.final_norm.variance_epsilon if hasattr(self.final_norm, "variance_epsilon") else self.final_norm.eps self._schema_cache = OrderedDict() # LRU of cached question prefixes self.max_schemas = 64 def _load_head(self, f): z = np.load(f) t = lambda k: torch.tensor(z[k], dtype=torch.float32, device=self.device) return {"U": t("U"), "V": t("V"), "b": t("b"), "c": t("c"), "T": float(z["T"]), "n_pos": int(z["n_pos"])} def _enc(self, s): return self.tok.encode(s, add_special_tokens=False) @torch.inference_mode() def _run(self, ids, cache, image=None): """image: (embeddings [n, d], grid_thw) for the image-pad tokens in ids, or None.""" if image is None: out = self.text(input_ids=torch.tensor([ids], device=self.device), past_key_values=cache, use_cache=True) return out.last_hidden_state[0] from vision_common import mrope_positions emb, grid = image x = torch.tensor([ids], device=self.device) h = self.text.embed_tokens(x) mask = (x == self.image_token_id)[..., None].expand_as(h) h = h.masked_scatter(mask, emb.to(h.dtype)) pos = torch.tensor(mrope_positions(ids, cache.get_seq_length(), grid, self.image_token_id), device=self.device)[:, None, :] out = self.text(inputs_embeds=h, position_ids=pos, past_key_values=cache, use_cache=True) return out.last_hidden_state[0] @torch.inference_mode() def _image(self, image, max_side): from vision_common import MAX_SIDE_LIMIT, MIN_SIDE, image_block, prepare_image if getattr(self.model.model, "visual", None) is None or self.image_token_id is None: raise ValueError("this model has no vision tower") if not MIN_SIDE <= int(max_side) <= MAX_SIDE_LIMIT: raise ValueError(f"image_max_side must be between {MIN_SIDE} and {MAX_SIDE_LIMIT}") patches, grid = prepare_image(image, int(max_side)) out = self.model.model.get_image_features(torch.tensor(patches, device=self.device), torch.tensor([grid], device=self.device)) emb = out.pooler_output if hasattr(out, "pooler_output") else out if isinstance(emb, (list, tuple)): emb = torch.cat(list(emb), dim=0) return image_block(self.tok, grid), (emb, grid) def _prefix(self, key, text): if key in self._schema_cache: self._schema_cache.move_to_end(key) else: while len(self._schema_cache) >= self.max_schemas: self._schema_cache.popitem(last=False) cache = DynamicCache(config=self.text.config) self._run(self._enc(text), cache) self._schema_cache[key] = cache return self._schema_cache[key] def _cap(self, state, reserved): limit = max(32, self.budget - reserved) ids = self._enc(state) if len(ids) <= limit: return state h = int(limit * 0.4) return self.tok.decode(ids[:h]) + " ... " + self.tok.decode(ids[-(limit - h):]) @torch.inference_mode() def hidden(self, state, questions, cache_schema=True, image=None, image_max_side=576): """Layer-56 residual at each trigger (used by ask(), and to re-fit the readout head on this backend).""" blocks, meta = [], [] for i, (name, spec) in enumerate(questions.items(), 1): b, labels, opts = _block(i, name, spec) blocks.append(b) meta.append((i, name, spec, labels, opts)) qtext = "QUESTIONS:\n" + "\n\n".join(blocks) trig_n = sum(len(self._enc(("" if k == 0 else "\n") + f"Q{i}:")) for k, (i, *_ ) in enumerate(meta)) reserved = len(self._enc(self.tail)) + trig_n + 4 img_ids, img = self._image(image, image_max_side) if image is not None else ([], None) if cache_schema: base = self._prefix(json.dumps(list(questions.items()), sort_keys=True), self.head + qtext + "\n\nSTATE:\n") body = self._cap(state, reserved) if img is None: seq = self._enc(body + self.tail) else: from vision_common import noul_reminder seq = img_ids + self._enc(("\n" + body if body else "") + noul_reminder(questions) + self.tail) else: base = self._prefix("__head__", self.head) reserved += len(self._enc("STATE:\n\n\n" + qtext)) if img is None: seq = self._enc("STATE:\n" + self._cap(state, reserved) + "\n\n" + qtext + self.tail) else: text = self._cap(state, reserved) seq = (self._enc("STATE:\n") + img_ids + self._enc(("\n" + text if text else "") + "\n\n" + qtext + self.tail)) pos = [] for k, (i, *_ ) in enumerate(meta): seq += self._enc(("" if k == 0 else "\n") + f"Q{i}:") pos.append(len(seq) - 1) h = self._run(seq, copy.deepcopy(base), img)[pos].to(torch.float32) return h, meta, len(seq) @torch.inference_mode() def ask(self, state, questions, cache_schema=True, image=None, image_max_side=576): h, meta, n = self.hidden(state, questions, cache_schema, image, image_max_side) hd = self.vision_head if image is not None else self.text_head pb = hd["c"][torch.clamp(torch.arange(len(meta), device=self.device), max=hd["n_pos"] - 1)] x = h + (h @ hd["U"]) @ hd["V"] + hd["b"] + pb x = x * torch.rsqrt((x * x).mean(-1, keepdim=True) + self.eps) * self.norm_w logits = (x @ self.W.T).cpu().numpy() res = {} for k, (i, name, spec, labels, opts) in enumerate(meta): z = logits[k][[VOCAB.index(l) for l in labels]] / hd["T"] p = np.exp(z - np.logaddexp.reduce(z)) j = int(p.argmax()) r = {"probs": {str(o): float(v) for o, v in zip(opts, p)}, "probs_list": [float(v) for v in p], "confidence": float(p[j])} r["value"] = float(p[0]) if spec["type"] == "noul" else opts[j] if spec["type"] == "score": r["expected_index"] = float((p * np.arange(len(p))).sum()) res[name] = r res["_tokens"] = n return res if __name__ == "__main__": import sys import time d = Decider(sys.argv[1] if len(sys.argv) > 1 else "dowands/neojev", head=sys.argv[2] if len(sys.argv) > 2 else None) qs = {"queue": {"type": "choice", "options": ["Billing and Payments", "Technical Support", "Sales"], "prompt": "Which team should handle this ticket?"}, "urgency": {"type": "score", "levels": ["low", "medium", "high"], "prompt": "How urgent is it?"}, "churn": {"type": "noul", "statement": "The customer is likely to leave."}} s = "Customer emailed twice about a failed refund for order #4417 and says they will dispute the charge tomorrow." d.ask(s, qs) torch.cuda.synchronize() t = time.perf_counter() out = d.ask(s, qs) torch.cuda.synchronize() print(json.dumps(out, indent=1)) print(f"{(time.perf_counter() - t) * 1000:.0f} ms")