lumierenoir commited on
Commit
f5f77d4
·
verified ·
1 Parent(s): ff3fb1b

base/: install section

Browse files
Files changed (1) hide show
  1. base/README.md +34 -0
base/README.md CHANGED
@@ -66,6 +66,40 @@ kev's *own* decision state already recovers the hate-speech signal with no exter
66
  which is why there is little left for the stitch to add. That holds across two different
67
  external adapters, so it looks like a property of kev's decision representation on this task.
68
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
69
  ## Using it
70
 
71
  Identical to the top-level card except the base — see its **Usage** section for the full guide
 
66
  which is why there is little left for the stitch to add. That holds across two different
67
  external adapters, so it looks like a property of kev's decision representation on this task.
68
 
69
+ ## Install
70
+
71
+ ```bash
72
+ pip install klev
73
+ ```
74
+
75
+ `klev` is the library: it carries the loader, the pointer readout, the record format and both
76
+ usage paths, so a consumer does not clone this repo to use a checkpoint. Two version floors
77
+ matter and pip does not know they are related to the architecture:
78
+
79
+ | | why |
80
+ |---|---|
81
+ | `transformers >= 5.5.0` | the `qwen3_5` and `gemma4` architectures do not exist before it. On 5.3.0 `AutoConfig` raises "Transformers does not recognize this architecture", which unsloth then **misreports** as a generic "not supported yet" ValueError — it reads like a missing model, not a version. |
82
+ | `unsloth >= 2026.9.11` | first release carrying gemma4 support. |
83
+
84
+ Install unsloth with `--no-deps` if you are on ROCm, so its CUDA `xformers` dependency does not
85
+ replace your ROCm torch build.
86
+
87
+ On AMD/ROCm also `source scripts/rocm_env.sh` from the repo (shipped in the wheel as
88
+ `klev.scripts`) **before** importing — it sets the arch override, `TORCH_ROCM_AOTRITON_ENABLE_BF16=0`,
89
+ `UNSLOTH_COMPILE_DISABLE=1` and the rest. All of it is load-bearing; see
90
+ `docs/m11-qwen35-08b.md`.
91
+
92
+ Two console scripts come with it:
93
+
94
+ ```bash
95
+ klev-system-one --ckpt lumierenoir/klev-0.8b --request my_request.json # no server
96
+ klev-serve --ckpt lumierenoir/klev-0.8b --port 8090 # optional HTTP
97
+ ```
98
+
99
+ `load_klev()` reads the base model from the checkpoint's own `adapter_config.json` and picks the
100
+ matching delimiters, so you do not have to pass `base=` or `preset=`. Pairing a checkpoint with
101
+ another model's weights would otherwise be silent nonsense rather than an error.
102
+
103
  ## Using it
104
 
105
  Identical to the top-level card except the base — see its **Usage** section for the full guide