| import sys |
|
|
| sys.path.insert(0, "/tmp/opencode/tl") |
| import importlib |
|
|
| import numpy as np |
|
|
| import export_full as E |
|
|
| GREEDY_FLOAT = [4612, 2619, 260, 647, 31844, 635, 393, 260, 1496, 2789, 3557, 21075, 31843, 1100, |
| 4838, 289, 826, 351, 600, 17360, 1086, 31843, 1591, 1124, 31844, 700, 2815, 260, |
| 1421, 3476, 288, 600, 2103, 31843, 1100, 393, 900, 6055, 291, 2598] |
|
|
|
|
| def q_grid(w, grid=13): |
| w = w.astype(np.float32) |
| best = None |
| for c in np.linspace(0.5, 1.0, grid): |
| s = np.abs(w).max(-1, keepdims=True) * c / 7.0 |
| s[s == 0] = 1e-12 |
| q = np.clip(np.round(w / s), -8, 7) |
| e = ((q * s - w) ** 2).sum(-1, keepdims=True) |
| if best is None: |
| best = (q, s, e) |
| else: |
| m = e < best[2] |
| best = (np.where(m, q, best[0]), np.where(m, s, best[1]), np.minimum(e, best[2])) |
| return best[0], best[1][..., 0] |
|
|
|
|
| def q_diffuse(w, grid=13): |
| q, s = q_grid(w) |
| q = q.copy().astype(np.float32) |
| w = w.astype(np.float32) |
| for i in range(w.shape[0]): |
| order = np.argsort(-np.abs(w[i])) |
| buf = w[i].copy() |
| for j in order: |
| v = np.clip(np.round(buf[j] / s[i]), -8, 7) |
| q[i, j] = v |
| err = buf[j] - v * s[i] |
| if j + 1 < len(buf): |
| buf = buf + 0.0 |
| q[i] = np.clip(np.round(w[i] / s[i]), -8, 7) |
| return q.astype(np.int8), s |
|
|
|
|
| def q_gptq(w, grid=13, damp=0.01): |
| """column-wise error feedback with an identity Hessian approximation (no calibration needed)""" |
| w = w.astype(np.float64) |
| q, s = q_grid(w) |
| sr = s.astype(np.float64) |
| H = np.eye(w.shape[1]) |
| for i in range(w.shape[0]): |
| W = w[i].copy() |
| for j in range(w.shape[1]): |
| qq = np.clip(np.round(W[j] / sr[i]), -8, 7) |
| q[i, j] = qq |
| err = (W[j] - qq * sr[i]) / H[j, j] |
| W = W - err * H[j] |
| return q.astype(np.int8), s |
|
|
|
|
| def run(tag, quant): |
| E.quantize = quant |
| for m in list(sys.modules): |
| if m in ("mirror3",): |
| del sys.modules[m] |
| import mirror3 as M |
|
|
| M.ATT_C, M.PKSH, M.SILU_SH, M.SILU_RND = 66572, 0, 11, False |
| M.build_tables() |
| st = M.State() |
| tok = 1 |
| out = [] |
| for pos in range(40): |
| lg = M.step(tok, pos, st) |
| tok = int(lg.argmax()) |
| out.append(tok) |
| n = 0 |
| for a, b in zip(out, GREEDY_FLOAT): |
| if a != b: |
| break |
| n += 1 |
| print(f"{tag:22s} coinciden {n}/40 tokens: {out[:4]}...{out[36:]}", flush=True) |
|
|
|
|
| run("grid (actual)", q_grid) |
| run("difusion de error", q_diffuse) |
| run("GPTQ identidad", q_gptq) |
|
|
