Commit Graph

Select branches

Hide Pull Requests

master

mega-moe-nvfp4

modelopt-nvfp4

nvidia-modelopt

Mono Color

• 0c77a88757 sync: latest Dockerfile + nvfp4_linear.py patch from B200 modelopt-nvfp4 biondizzle 2026-05-14 16:47:27 +00:00
• f2656dcf6d sync B200 deployment files: Dockerfile, docker-compose, patches biondizzle 2026-05-14 14:12:52 +00:00
• f08bcd456b tweax n shit mega-moe-nvfp4 biondizzle 2026-05-12 23:16:33 +00:00
• 2bdda36bb7 fucken aye biondizzle 2026-05-12 22:33:55 +00:00
• fa825c16b9 fucken ay biondizzle 2026-05-12 21:58:08 +00:00
• dbf1d11f9f ayee biondizzle 2026-05-12 21:52:48 +00:00
• ef3edb3481 ba fongol again4 biondizzle 2026-05-12 21:49:43 +00:00
• b74dc7121a ba fongol again3 biondizzle 2026-05-12 21:48:00 +00:00
• 7bbbdbcc79 ba fongol again biondizzle 2026-05-12 21:47:37 +00:00
• 2e674f87c1 ba fongol again biondizzle 2026-05-12 21:47:14 +00:00
• 5d127d8294 ba fongol again biondizzle 2026-05-12 21:47:01 +00:00
• 52cf3f2e25 ba fongol again biondizzle 2026-05-12 21:46:42 +00:00
• 02decb486e ba fongol biondizzle 2026-05-12 21:34:08 +00:00
• 48f1f9dc5e clanker nonsense again biondizzle 2026-05-12 21:30:36 +00:00
• 5cabc1f7d9 clanker nonsense again biondizzle 2026-05-12 21:29:59 +00:00
• 25a2d4e6ad clanker nonsense biondizzle 2026-05-12 21:28:50 +00:00
• d88ea9842b fix: add missing staging_kernel.py to Dockerfile — BF16→E2M1+UE4M3 quantization was never in container biondizzle 2026-05-12 21:21:24 +00:00
• 91d7d9bad7 fucken a biondizzle 2026-05-12 21:18:48 +00:00
• d68e113af1 remove spammy shit biondizzle 2026-05-12 20:57:04 +00:00
• f0652693a6 dangit again biondizzle 2026-05-12 19:13:01 +00:00
• 054792c84e dangit biondizzle 2026-05-12 18:42:39 +00:00
• de055b1e77 syupid clankers biondizzle 2026-05-12 18:26:37 +00:00
• 307574bc91 test: signal alarm timeout for kernel hang biondizzle 2026-05-12 15:14:39 +00:00
• fcd6de0a60 test: simplify SF fill to avoid shape mismatch biondizzle 2026-05-12 15:13:16 +00:00
• d4c557fddc test: fix float8 randn + SF int32 packing biondizzle 2026-05-12 15:12:35 +00:00
• 28afc2406b test: add random FP4 data and kernel timeout biondizzle 2026-05-12 15:11:41 +00:00
• 787d427847 test: fix NVFP4 mega_moe test dimensions for SMEM alignment biondizzle 2026-05-12 15:07:35 +00:00
• 8737fd57c0 remove crap biondizzle 2026-05-12 14:53:42 +00:00
• 52c3aefe73 bump cache busters to 33 for debug build biondizzle 2026-05-12 13:10:37 +00:00
• ca1d306890 fix: use torch.int8 for packed FP4 tensors (kPackedFP4=kInt8, not uint8) biondizzle 2026-05-12 12:23:43 +00:00
• b8f95ffad3 docker: add OMP_NUM_THREADS=64, remove --tool initcheck, mount cubin cache biondizzle 2026-05-12 11:15:06 +00:00
• 5840291ea3 fix staging kernel packed_k_mask double-count biondizzle 2026-05-12 08:08:24 +00:00
• 5ea5b579c3 Trim banner, no code changes biondizzle 2026-05-12 07:24:36 +00:00
• 74af9984f6 Bug fixes: UE4M3 scale conversion, staging kernel SF/E2M1 packing, wo_a UE4M3, README overhaul biondizzle 2026-05-12 05:52:30 +00:00
• a36bf47f11 fix: use tl.split instead of indexing for E2M1 pair packing biondizzle 2026-05-11 22:39:38 +00:00
• 27dbf2850f fix: replace nested tl.where with sum-of-comparisons for E2M1 quantization biondizzle 2026-05-11 22:23:05 +00:00
• 3d1f3de190 fix: syntax error — move triton imports before docstring, remove orphan @triton.jit biondizzle 2026-05-11 22:08:50 +00:00
• 79d866995f bump cache buster 32 for packed FP4 mxf4nvf4 fix biondizzle 2026-05-11 21:59:56 +00:00
• c85b84b0fe fix: staging kernel outputs unpacked E2M1 (1 byte/element, not packed 2/byte) biondizzle 2026-05-11 21:29:33 +00:00
• 01cfd02759 fix: same reshape fix in main patch file biondizzle 2026-05-11 21:05:54 +00:00
• 076d325c97 fix: use reshape instead of risky [0::2] slicing for E2M1 packing biondizzle 2026-05-11 21:04:53 +00:00
• 8dc917c498 fix: topk_weights_out store missing topk_offsets multiplier biondizzle 2026-05-11 21:02:19 +00:00
• 17ba5a9d7b bump cache buster 30 for FP4 staging + DeepGEMM FP4 activations biondizzle 2026-05-11 20:30:14 +00:00
• 7a4403fa98 feat: FP4 staging kernel - BF16 → E2M1 packed + UE4M3 block16 scales biondizzle 2026-05-11 20:29:36 +00:00
• 0fd2d4f078 diag: add weight_scale uint8 histogram to verify E8M0 vs E4M3 format biondizzle 2026-05-11 19:55:41 +00:00
• 50a945bde4 bump cache buster 29 biondizzle 2026-05-11 19:51:48 +00:00
• 48b905406a diag: add CUDA sync after mega_moe finalize + forward to catch errors biondizzle 2026-05-11 19:51:44 +00:00
• 35f6b66678 fix: UE8M0 reinterpret in DeepGEMM fold_global_scale + bump cache biondizzle 2026-05-11 19:40:08 +00:00
• f32d6b5b48 bump cache buster to 27 biondizzle 2026-05-11 19:26:21 +00:00
• cd24182e36 diag: add NaN/Inf + FP8-dtype checks after NVFP4 dequant biondizzle 2026-05-11 19:26:12 +00:00
• 8ae2214bad fix: reorder Dockerfile ARG before COPY for proper cache busting biondizzle 2026-05-11 18:48:07 +00:00
• c4891e9ee2 fix: manual FP32→UE4M3 quant in Triton staging kernel biondizzle 2026-05-11 16:38:49 +00:00
• 436109081c bump cache buster to 24 biondizzle 2026-05-11 16:12:56 +00:00
• 5faf9916eb fix: UE4M3 activation scales + group_size=16 for NVFP4 mega_moe biondizzle 2026-05-11 16:12:36 +00:00
• 220649c188 docs: CORRECTED — mxf4nvf4 IS supported on sm_100a (B200) biondizzle 2026-05-11 14:24:13 +00:00
• cfead0012d docs: comprehensive README update through build 22 biondizzle 2026-05-11 13:53:41 +00:00
• 8cb23bdb78 fix: import NVFP4 SymmBuffer from deep_gemm.mega biondizzle 2026-05-11 08:05:50 +00:00
• ff579c9767 fix: use NVFP4 SymmBuffer (2x SF size for group_size=16) biondizzle 2026-05-11 07:49:11 +00:00
• 1da40c53da fix: add patch cache buster to Dockerfile biondizzle 2026-05-11 07:19:10 +00:00
• b532742530 debug: add shape/dtype logging to finalize_weights biondizzle 2026-05-11 07:13:44 +00:00
• b1cf4232ee feat: wire DeepGEMM NVFP4 mega_moe kernel into vLLM patch biondizzle 2026-05-11 06:22:11 +00:00
• a2e9b5f17f fix: add --enable-expert-parallel to compose command biondizzle 2026-05-11 06:15:11 +00:00
• c8564caf9d fix: patch vLLM deepseek_v4.py directly in image biondizzle 2026-05-11 06:09:40 +00:00
• 7c8c6cd67f fix: add PYTHONPATH for deep_gemm import biondizzle 2026-05-11 06:06:52 +00:00
• cffb373759 fix: symlink NVRTC lib into cuda/lib64 for linker biondizzle 2026-05-11 06:04:24 +00:00
• 983ba02c5b fix: add CUDA/NVRTC lib paths to Dockerfile biondizzle 2026-05-11 06:02:13 +00:00
• f0471ed1c2 fix: correct CR URL to atl.vultrcr.com biondizzle 2026-05-11 05:59:06 +00:00
• c234190a80 feat: add Dockerfile + build/push script for NVFP4 container biondizzle 2026-05-11 05:57:49 +00:00
• e963325b61 WIP: MegaMoE NVFP4 kernel + diagnostics biondizzle 2026-05-11 05:19:49 +00:00
• 7e2f219259 fix: banner uses _os instead of os (not yet imported) biondizzle 2026-05-11 04:57:24 +00:00
• cf54b4755a fix CRITICAL #7: UE8M0 block scale misinterpreted as E4M3 biondizzle 2026-05-11 04:37:33 +00:00
• 7febeaeb71 README: document bugs #5 (input_scale) and #6 (fused_skip_regex), add version banner section, update status biondizzle 2026-05-11 04:28:38 +00:00
• 26aaaba4a2 Add version banner to patch — prints commit, arch, bugs fixed at startup biondizzle 2026-05-11 04:28:10 +00:00
• 67f9086a26 Fix critical dequantization bug: remove input_scale from weight dequant biondizzle 2026-05-11 02:23:18 +00:00
• 02b8ea536f Update MEMORY.md and memory files with vLLM NVFP4 serving progress biondizzle 2026-05-11 02:02:14 +00:00
• 653e2d7a50 vLLM NVFP4 serving: full end-to-end pipeline working biondizzle 2026-05-11 02:01:46 +00:00
• db16be8e5d S11: Fixed substr mapping, stacking, suffix, and o_a_proj - loads weights but attention forward uses FP8 einsum incompatible with NVFP4 biondizzle 2026-05-10 17:45:53 +00:00
• 6fd03a0aa0 vLLM serving: patched deepseek_v4.py, disabled mega_moe, updated docs biondizzle 2026-05-10 16:14:17 +00:00
• d88793dee6 Add vllm weight mapper patch and docker-compose biondizzle 2026-05-10 09:33:48 +00:00
• 30608e3834 Config patches: document modelopt↔vllm gaps with NVIDIA reference biondizzle 2026-05-10 08:59:28 +00:00
• 0d74b97fb2 Config patches doc + compress_ratios runtime patch in serve script biondizzle 2026-05-10 08:23:11 +00:00
• f65d4ab99f Run 11 SUCCESS: 881GB NVFP4 exported, add vLLM serve script biondizzle 2026-05-10 07:54:34 +00:00
• eb80bd6f80 README + memory: Run 10 result (export crash in get_weight_scaling_factor), Run 11 running biondizzle 2026-05-09 23:00:17 +00:00
• 07cd50e823 8 patches covering full export chain — no more whack-a-mole biondizzle 2026-05-09 22:50:58 +00:00
• efc111a11f Add Patch 4+5: get_weight_scaling_factor and get_weight_scaling_factor_2 CPU safety biondizzle 2026-05-09 22:43:48 +00:00
• ce9056d259 README overhaul: reflect current architecture (hf_main, run history through Run 10) biondizzle 2026-05-09 16:09:09 +00:00
• 5a72da7193 Fix: apply hf_ptq __main__ post-parse conversions (dataset split, calib_size int list) biondizzle 2026-05-09 15:58:36 +00:00
• 8612914169 Update run history: Runs 7-8, Run 9 running on a300302 biondizzle 2026-05-09 15:00:23 +00:00
• a300302486 Fix: use hf_ptq.py arg names (--pyt_ckpt_path, --qformat, --inference_tensor_parallel) biondizzle 2026-05-09 14:57:28 +00:00
• 1a36a655ea Fix: use full argparse flag names (--calib_size, --kv_cache_qformat) biondizzle 2026-05-09 14:54:51 +00:00
• b2849a8944 Fundamental rewrite: call hf_main() instead of rewriting the pipeline biondizzle 2026-05-09 14:52:02 +00:00
• a70593d886 Update run history: Run 6 (dataloader crash), Run 7 running on 25b4d8d biondizzle 2026-05-09 13:40:00 +00:00
• 25b4d8da06 Fix: add missing args for make_calib_dataloader (dataset, calib_with_images, auto_quantize, specdec) biondizzle 2026-05-09 13:37:24 +00:00
• d1e15178b2 Update run history: Runs 4-5 (import bugs), Run 6 running on 6c1bff6 biondizzle 2026-05-09 09:29:20 +00:00
• 6c1bff6997 Clean rewrite: verified all imports against runtime, removed dead code biondizzle 2026-05-09 09:26:23 +00:00
• 86dd8df302 Fix: KV_QUANT_CFG_CHOICES is in hf_ptq, not mtq biondizzle 2026-05-09 09:17:12 +00:00
• 99f861f48a Update README and memory: Run 3 OOM crash, Run 4 running on f9bbef8 biondizzle 2026-05-09 08:10:04 +00:00
• f9bbef8e91 Fix: patch load_calib_amax instead of amax property setter (can't patch readonly descriptor) biondizzle 2026-05-09 08:04:03 +00:00
• 94179ed9d0 Fix typo: store_only → store_true biondizzle 2026-05-09 08:02:09 +00:00
• 03c10ab3b6 Fix model loading: use modelopt get_model() instead of raw AutoModelForCausalLM biondizzle 2026-05-09 08:00:50 +00:00