bitsandbytes-cpu-fork 0.50.2.dev2__py3-none-win_amd64.whl → 0.50.2.dev3__py3-none-win_amd64.whl

{bitsandbytes_cpu_fork-0.50.2.dev2.dist-info → bitsandbytes_cpu_fork-0.50.2.dev3.dist-info}/METADATA

@@ -1,7 +1,7 @@
 Metadata-Version: 2.4
 Name: bitsandbytes-cpu-fork
-Version: 0.50.2.dev2
-Summary: bitsandbytes with a pure-CPU training backend (AVX2 / NEON): 8-bit optimizers, fused GDN, gemm_8bit, disk balancer. No GPU required.
+Version: 0.50.2.dev3
+Summary: bitsandbytes with a pure-CPU training backend (AVX2 / NEON): 8-bit and 4-bit optimizers, fused GDN, gemm_8bit, disk balancer. No GPU required.
 License-Expression: MIT
 Project-URL: repository, https://github.com/Ganyvnanoaa09-0906/bitsandbytes-CPU
 Project-URL: Upstream bitsandbytes, https://github.com/bitsandbytes-foundation/bitsandbytes
@@ -31,7 +31,7 @@ Dynamic: license-file
 
 > **This is the CPU fork** — `bitsandbytes-cpu-fork`, not upstream `bitsandbytes`.
 > It replaces the CUDA kernels with a pure-CPU backend (AVX2 on x86-64, NEON on ARM64):
-> 8-bit optimizers, fused Gated DeltaNet, `gemm_8bit`, block-wise 4/8-bit quantisation.
+> 8/4-bit optimizers, fused Gated DeltaNet, `gemm_8bit`, block-wise 4/8-bit quantisation.
 > No GPU, no CUDA toolkit, no NVIDIA driver. Full guide:
 > [`docs_cpu/QUICKSTART.md`](docs_cpu/QUICKSTART.md) (中文) ·
 > [`docs_cpu/QUICKSTART_EN.md`](docs_cpu/QUICKSTART_EN.md).
@@ -354,6 +354,7 @@ Building from source instead (any platform) is described in
 | **Gated DeltaNet fwd/bwd** | `gdn_fwd_cpu` / `gdn_bwd_cpu` (csrc/cpu_gdn.cpp) | Fused Gated DeltaNet (Qwen3-Next/3.5 linear attention) — ~35× vs per-step Python loop; chunk-checkpointed backward |
 | **Fused 8-bit blockwise dequant GEMM** | `cgemm_8bit_inference_cpu_fp32` (`bitsandbytes::gemm_8bit`, `bnb.functional.fused_dequant_linear_8bit`) | `out = A @ dequant8(B)` with weights kept uint8 (¼ DRAM traffic), no fp32 temporary |
 | **Fused 8-bit optimizer** | `coptimizer_update_8bit_blockwise_cpu` (`bnb.optim.AdamW8bit` etc.) | Single-pass dequant→update→requant; optimizer state memory ≈ 1/3.8 of fp32 |
+| **Fused 4-bit optimizer** | `coptimizer_update_4bit_blockwise_cpu` (`bnb.optim.AdamW4bit`) | Same single-pass shape as the 8-bit one at half the state: **1.031 vs 2.062 B/param** at 1M params, and **0.658 ms vs 1.682 ms** per step (AVX2). On a real 30.49M model the state drops 232.6 MB to 29.5 MB. Trains to the same place: loss curves track fp32 within 0.27% over 120 steps, state codes identical to an independent numpy reference over 50 |
 | **Blockwise 8/4-bit quant & dequant** | `cquantize_blockwise_cpu_*` / `cdequantize_blockwise_cpu_*` | AVX2 LUT quantize + dequantize; NF4/FP4 kernels with NEON/AVX2 paths |
 | **4-bit inference GEMV** | `cgemv_4bit_inference_cpu_*` | Fused 4-bit dequant GEMV/GEMM for AVX2 machines (symbol alias fixed) |
 | **GDN runtime patch** | `bitsandbytes/gdn_cpu.py` | `patch_transformers()` / `patch_fla()` — route transformers GDN slow paths to the fused kernel |

{bitsandbytes_cpu_fork-0.50.2.dev2.dist-info → bitsandbytes_cpu_fork-0.50.2.dev3.dist-info}/RECORD

@@ -1,5 +1,5 @@
-bitsandbytes_cpu_cli.py,sha256=PVpzzjB_jv3zHMudm4U14VmWMf95EiWf2lyMT54Mxl4,47770
-bitsandbytes/__init__.py,sha256=akyoGIUi8Gbe9H46rYGJ7JTL72KryOPGIs3zwdnwnUM,2370
+bitsandbytes_cpu_cli.py,sha256=rSRi1aba5Ztr5HGrC0is8wG_hWxJx2W8CD9FD-rm96U,50540
+bitsandbytes/__init__.py,sha256=3LUcCQqCYospyzYIk_d1kw6Y0jsfA9QyOCXi1IOp0jM,2370
 bitsandbytes/__main__.py,sha256=_ObgXmW-NG395jj_oP86gSOw2LgqMtrGBmpjZOiRcuY,94
 bitsandbytes/_ops.py,sha256=G6jalkQVarTc0riBO7EEPYfC4u1iEAX-JGmlnkzxB8g,21166
 bitsandbytes/cextension.py,sha256=7TjyZFk1w1CTI4SWZDrGMPnVlMVXtpMvDZgAz_e9228,17579
@@ -55,9 +55,9 @@ bitsandbytes/optim/lion.py,sha256=WmRuaeVntkjkKvIivVh7t3Ac-nqKWFxtZ3NShQc3rJM,85
 bitsandbytes/optim/optimizer.py,sha256=pS8X47Bo3sgkDDIkGEp8C6stE-ZM_Hx1DvFWExyg9MI,28883
 bitsandbytes/optim/rmsprop.py,sha256=wJel4HgLjCPEHjcB5vPFZF6HleKD4vh1ILYt0dPR0is,6304
 bitsandbytes/optim/sgd.py,sha256=UYydMLYpBnW3nj4tnIq5FYjkhOLXz8YBaISKHQ3go6Q,5068
-bitsandbytes_cpu_fork-0.50.2.dev2.dist-info/licenses/LICENSE,sha256=8IqUv7C9wdfxy1m62vD4xVGxDbhxGFYi2_H9A2r_qcs,1107
-bitsandbytes_cpu_fork-0.50.2.dev2.dist-info/METADATA,sha256=eG0V2Kncyua9i8M1fCL-TLvPXAQZ2mzisUESb3PvyVQ,18241
-bitsandbytes_cpu_fork-0.50.2.dev2.dist-info/WHEEL,sha256=5s7VO_HUPfjwcO4LINqxGe8JSFZAGgaurQKPGA4cx_g,97
-bitsandbytes_cpu_fork-0.50.2.dev2.dist-info/entry_points.txt,sha256=GqLu5Rm-j-9_CvhqX0jVecyJShA1XGmxRoPEiibJpcI,99
-bitsandbytes_cpu_fork-0.50.2.dev2.dist-info/top_level.txt,sha256=l9ip7vjQq1JV6CB_ZGntWtjfxpkH119NDxupnJho9k0,34
-bitsandbytes_cpu_fork-0.50.2.dev2.dist-info/RECORD,,
+bitsandbytes_cpu_fork-0.50.2.dev3.dist-info/licenses/LICENSE,sha256=8IqUv7C9wdfxy1m62vD4xVGxDbhxGFYi2_H9A2r_qcs,1107
+bitsandbytes_cpu_fork-0.50.2.dev3.dist-info/METADATA,sha256=aHjxngO9F2nY9FuSGr3Qf5gdOuRaQ8kbdi-m6xmwfAg,18699
+bitsandbytes_cpu_fork-0.50.2.dev3.dist-info/WHEEL,sha256=5s7VO_HUPfjwcO4LINqxGe8JSFZAGgaurQKPGA4cx_g,97
+bitsandbytes_cpu_fork-0.50.2.dev3.dist-info/entry_points.txt,sha256=GqLu5Rm-j-9_CvhqX0jVecyJShA1XGmxRoPEiibJpcI,99
+bitsandbytes_cpu_fork-0.50.2.dev3.dist-info/top_level.txt,sha256=l9ip7vjQq1JV6CB_ZGntWtjfxpkH119NDxupnJho9k0,34
+bitsandbytes_cpu_fork-0.50.2.dev3.dist-info/RECORD,,