Commits · 5223199e03ac3729eb60043a1ef57156c8af1bc9 · OpenDAS / vllm_cscc · GitLab

07 Aug, 2024 3 commits
- [Bugfix][FP8] Fix dynamic FP8 Marlin quantization (#7219) · 5223199e
  Michael Goin authored Aug 07, 2024
  
  5223199e
- [Misc] Refactor linear layer weight loading; introduce `BasevLLMParameter` and... · 0f7052bc
  Dipika Sikka authored Aug 07, 2024
```
[Misc] Refactor linear layer weight loading; introduce `BasevLLMParameter` and `weight_loader_v2` (#5874)
```
  0f7052bc
- [Bugfix] Fix GPTQ and GPTQ Marlin CPU Offloading (#7225) · f9a56006
  Michael Goin authored Aug 06, 2024
  
  f9a56006
05 Aug, 2024 1 commit
- [Core] Support loading GGUF model (#5191) · 360bd67c
  Isotr0py authored Aug 06, 2024
```
Co-authored-by: Michael Goin <michael@neuralmagic.com>
```
  360bd67c
02 Aug, 2024 1 commit
- [Misc] Disambiguate quantized types via a new ScalarType (#6396) · a8d604ca
  Lucas Wilkinson authored Aug 02, 2024
  
  a8d604ca
31 Jul, 2024 1 commit
- Support W4A8 quantization for vllm (#5218) · 6512937d
  HandH1998 authored Jul 31, 2024
  
  6512937d
30 Jul, 2024 1 commit
- [Kernel] Remove scaled_fp8_quant kernel padding footgun (#6842) · d7a299ed
  Tyler Michael Smith authored Jul 30, 2024
  
  d7a299ed
29 Jul, 2024 1 commit
- [Bugfix] Allow vllm to still work if triton is not installed. (#6786) · 9a7e2d05
  Thomas Parnell authored Jul 29, 2024
```
Signed-off-by: Thomas Parnell <tpa@zurich.ibm.com>
```
  9a7e2d05
28 Jul, 2024 1 commit
- [Misc] Pass cutlass_fp8_supported correctly in fbgemm_fp8 (#6871) · 3eeb148f
  Elsa Granger authored Jul 28, 2024
  
  3eeb148f
27 Jul, 2024 1 commit
- [Kernel] Increase precision of GPTQ/AWQ Marlin kernel (#6795) · 75acdaa4
  Alexander Matveev authored Jul 27, 2024
  
  75acdaa4
25 Jul, 2024 4 commits
- [Bugfix] Fix empty (nullptr) channelwise scales when loading wNa16 using... · cd7edc4e
  Lucas Wilkinson authored Jul 25, 2024
```
[Bugfix] Fix empty (nullptr) channelwise  scales when loading wNa16 using compressed tensors (#6798)
```
  cd7edc4e
- [Bugfix] Fix `kv_cache_dtype=fp8` without scales for FP8 checkpoints (#6761) · 65b1f121
  Michael Goin authored Jul 25, 2024
  
  65b1f121
- [ Misc ] `fp8-marlin` channelwise via `compressed-tensors` (#6524) · 889da130
  Robert Shaw authored Jul 25, 2024
```
Co-authored-by: mgoin <michael@neuralmagic.com>
```
  889da130
- [Bugfix] Fix awq_marlin and gptq_marlin flags (#6745) · 0310029a
  Alexander Matveev authored Jul 25, 2024
  
  0310029a
23 Jul, 2024 4 commits
- [bitsandbytes]: support read bnb pre-quantized model (#5753) · 87525fab
  dongmao zhang authored Jul 23, 2024
```
Co-authored-by: Michael Goin <michael@neuralmagic.com>
```
  87525fab
- [Misc] Add ignored layers for `fp8` quantization (#6657) · 0eb0757b
  Michael Goin authored Jul 23, 2024
  
  0eb0757b
- [Misc] Support FP8 kv cache scales from compressed-tensors (#6528) · 9e0b558a
  Michael Goin authored Jul 23, 2024
  
  9e0b558a
- [Bugfix] Fix null `modules_to_not_convert` in FBGEMM Fp8 quantization (#6665) · c5e83309
  Cheng Li authored Jul 22, 2024
  
  c5e83309
21 Jul, 2024 1 commit
- [Kernel][Core] Add AWQ support to the Marlin kernel (#6612) · 396d92d5
  Alexander Matveev authored Jul 21, 2024
  
  396d92d5
20 Jul, 2024 4 commits
- [ Bugfix ] Fix AutoFP8 fp8 marlin (#6609) · 082ecd80
  Robert Shaw authored Jul 20, 2024
  
  082ecd80
- [Misc] Fix input_scale typing in w8a8_utils.py (#6579) · f952bbc8
  Michael Goin authored Jul 20, 2024
  
  f952bbc8
- [ Kernel ] Enable `fp8-marlin` for `fbgemm-fp8` models (#6606) · 9364f74e
  Robert Shaw authored Jul 20, 2024
  
  9364f74e
- [ Misc ] `fbgemm` checkpoints (#6559) · 683e3cb9
  Robert Shaw authored Jul 20, 2024
  
  683e3cb9
19 Jul, 2024 2 commits
- [ Kernel ] Enable Dynamic Per Token `fp8` (#6547) · 4cc24f01
  Robert Shaw authored Jul 19, 2024
  
  4cc24f01
- [ Misc ] non-uniform quantization via `compressed-tensors` for `Llama` (#6515) · dbe55885
  Robert Shaw authored Jul 18, 2024
  
  dbe55885
18 Jul, 2024 3 commits
- [Kernel] Implement fallback for FP8 channelwise using torch._scaled_mm (#6552) · 4ffffccb
  Tyler Michael Smith authored Jul 18, 2024
  
  4ffffccb
- [ Misc ] Improve Min Capability Checking in `compressed-tensors` (#6522) · 58ca6632
  Robert Shaw authored Jul 18, 2024
  
  58ca6632
- [ Kernel ] Fp8 Channelwise Weight Support (#6487) · 18fecc35
  Robert Shaw authored Jul 17, 2024
  
  18fecc35
16 Jul, 2024 1 commit
- [Kernel][Attention] Separate `Attention.kv_scale` into `k_scale` and `v_scale` (#6081) · 978aed53
  Michael Goin authored Jul 16, 2024
  
  978aed53
15 Jul, 2024 1 commit
- [Kernel] Use CUTLASS kernels for the FP8 layers with Bias (#6270) · c8fd97f2
  Tyler Michael Smith authored Jul 15, 2024
  
  c8fd97f2
14 Jul, 2024 1 commit
- [ Misc ] Apply MoE Refactor to Deepseekv2 To Support Fp8 (#6417) · fb6af8bc
  Robert Shaw authored Jul 13, 2024
  
  fb6af8bc
13 Jul, 2024 1 commit
- [ Misc ] More Cleanup of Marlin (#6359) · babf52da
  Robert Shaw authored Jul 13, 2024
```
Co-authored-by: Robert Shaw <rshaw@neuralmagic.com>
```
  babf52da
12 Jul, 2024 1 commit
- [ Misc ] Support Models With Bias in `compressed-tensors` integration (#6356) · aea19f09
  Robert Shaw authored Jul 12, 2024
  
  aea19f09
11 Jul, 2024 1 commit
- [ Misc ] Refactor Marlin Python Utilities (#6082) · b675069d
  Robert Shaw authored Jul 11, 2024
```
Co-authored-by: Robert Shaw <rshaw@neuralmagic.com>
```
  b675069d
07 Jul, 2024 1 commit
- [ Misc ] Support Fp8 via `llm-compressor` (#6110) · abfe705a
  Robert Shaw authored Jul 07, 2024
```
Co-authored-by: Robert Shaw <rshaw@neuralmagic>
```
  abfe705a
03 Jul, 2024 3 commits
- [ Misc ] Clean Up `CompressedTensorsW8A8` (#6113) · 62963d12
  Robert Shaw authored Jul 03, 2024
  
  62963d12
- [Kernel] Expand FP8 support to Ampere GPUs using FP8 Marlin (#5975) · 47f0954a
  Michael Goin authored Jul 03, 2024
  
  47f0954a
- [hardware][misc] introduce platform abstraction (#6080) · 482045ee
  youkaichao authored Jul 02, 2024
  
  482045ee
02 Jul, 2024 2 commits
- [CORE] Quantized lm-head Framework (#4442) · ee93f4f9
  Qubitium-ModelCloud authored Jul 03, 2024
```
Co-authored-by: Robert Shaw <rshaw@neuralmagic.com>
Co-authored-by: ZX <zx@lbx.dev>
```
  ee93f4f9
- [ Misc ] Refactor MoE to isolate Fp8 From Mixtral (#5970) · 7c008c51
  Robert Shaw authored Jul 02, 2024
```
Co-authored-by: Robert Shaw <rshaw@neuralmagic>
Co-authored-by: Michael Goin <michael@neuralmagic.com>
```
  7c008c51