mirror of
https://github.com/elder-plinius/OBLITERATUS.git
synced 2026-08-17 16:37:30 +02:00
Brian Bell
62b006f6c4
Point obliteratus at an FP8 or NVFP4 checkpoint and it just works: the loader detects the format from config.json + safetensors metadata (no weight loads), dequantizes shard-by-shard to a temporary BF16 copy, then runs the normal float pipeline and saves BF16. Supported layouts: - FP8 DeepSeek-style block-wise (weight_scale_inv + weight_block_size) - FP8 per-channel / per-tensor (compressed-tensors, ModelOpt) - NVFP4 ModelOpt (uint8 nibbles + FP8 group scales + FP32 global), including MIXED_PRECISION checkpoints (FP8 mixer + NVFP4 experts) - NVFP4 compressed-tensors (reciprocal scales) Design: - New pure-torch obliteratus/models/quant_dequant.py; no new deps. NVFP4 unpack uses torch.float4_e2m1fn_x2 when a runtime probe proves it works, else a chunked nibble LUT (bounds transient int64 index memory; a naive implementation OOMed at 96GB on a 30B model). - Scale keys are dropped only when their base weight exists in the same shard, so legitimate params ending in _scale (logit_scale et al.) survive. - Unsupported schemes (fbgemm, quanto, W4A4, ...) fail loudly at load, naming the scheme. - Surgery guards: float8 or packed uint8 reaching _dequantize_weight or any fused-MoE path raises RuntimeError instead of silently upcasting (bitsandbytes quant_state params are explicitly excluded). - Save path strips quantization metadata and logs that output is BF16; re-quantization for serving is out of scope (llm-compressor/modelopt). - CLI: new --trust-remote-code flag; help text documents auto-detection. Validated end-to-end on 1x A100-80GB (see PR description): Nemotron-3-Nano-Omni-30B NVFP4 (mixed) and FP8, Qwen3-8B-FP8 (block-wise) vs Qwen3-8B BF16 baseline (perplexity 4.23 vs 4.33).
56 lines
1.8 KiB
Python
56 lines
1.8 KiB
Python
"""Smoke test: load a quantized checkpoint through the OBLITERATUS loader.
|
|
|
|
Verifies: detection, dequantization to BF16, no quantized tensors left in
|
|
the model, config stripped of quantization_config, and a short generation.
|
|
"""
|
|
|
|
import sys
|
|
|
|
import torch
|
|
|
|
from obliteratus.models import quant_dequant as qd
|
|
from obliteratus.models.loader import load_model
|
|
|
|
repo = sys.argv[1]
|
|
|
|
det = qd.detect_quant_scheme(repo)
|
|
print(f"[detection] {det.scheme.value} (inverse={det.scale_is_inverse}, group={det.group_size})")
|
|
assert det.scheme not in (qd.QuantScheme.NONE, qd.QuantScheme.UNSUPPORTED), det
|
|
|
|
handle = load_model(
|
|
repo,
|
|
task="causal_lm",
|
|
device="auto",
|
|
dtype="bfloat16",
|
|
trust_remote_code=True,
|
|
)
|
|
model = handle.model
|
|
print(f"[load] ok — scheme tag: {getattr(model, '_obliteratus_dequantized_scheme', None)}")
|
|
|
|
bad = []
|
|
n_params = 0
|
|
for name, p in model.named_parameters():
|
|
n_params += 1
|
|
if p.dtype in qd.FP8_DTYPES or p.dtype == torch.uint8:
|
|
bad.append((name, str(p.dtype)))
|
|
if torch.isnan(p.data).any().item() if p.data.is_floating_point() else False:
|
|
bad.append((name, "NaN"))
|
|
print(f"[check] {n_params} params, quantized/NaN leftovers: {bad[:10] or 'NONE'}")
|
|
assert not bad, bad
|
|
|
|
assert getattr(model.config, "quantization_config", None) is None, "quantization_config survived"
|
|
|
|
tok = handle.tokenizer
|
|
prompt = "The capital of France is"
|
|
try:
|
|
inputs = tok(prompt, return_tensors="pt").to(model.device)
|
|
except Exception:
|
|
# Omni processors: fall back to bare tokenizer encode
|
|
ids = tok.encode(prompt, return_tensors="pt").to(model.device)
|
|
inputs = {"input_ids": ids}
|
|
with torch.no_grad():
|
|
out = model.generate(**inputs, max_new_tokens=16, do_sample=False)
|
|
text = tok.decode(out[0][-16:] if out.dim() > 1 else out[-16:], skip_special_tokens=True)
|
|
print(f"[generate] {text!r}")
|
|
print("SMOKE_LOAD_OK")
|