Create quantize.yaml
Browse files- quantize.yaml +49 -0
quantize.yaml
ADDED
|
@@ -0,0 +1,49 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
test_stage:
|
| 2 |
+
obcq_modifiers:
|
| 3 |
+
LogarithmicEqualizationModifier:
|
| 4 |
+
mappings: [
|
| 5 |
+
[["re:.*q_proj", "re:.*k_proj", "re:.*v_proj"], "re:.*input_layernorm"],
|
| 6 |
+
[["re:.*gate_proj", "re:.*up_proj"], "re:.*post_attention_layernorm"],
|
| 7 |
+
]
|
| 8 |
+
QuantizationModifier:
|
| 9 |
+
ignore:
|
| 10 |
+
# These operations don't make sense to quantize
|
| 11 |
+
- LlamaRotaryEmbedding
|
| 12 |
+
- LlamaRMSNorm
|
| 13 |
+
- SiLUActivation
|
| 14 |
+
- MatMulOutput_QK
|
| 15 |
+
- MatMulOutput_PV
|
| 16 |
+
# Skip quantizing the layers with the most sensitive activations
|
| 17 |
+
- model.layers.21.mlp.down_proj
|
| 18 |
+
- model.layers.7.mlp.down_proj
|
| 19 |
+
- model.layers.2.mlp.down_proj
|
| 20 |
+
- model.layers.8.self_attn.q_proj
|
| 21 |
+
- model.layers.8.self_attn.k_proj
|
| 22 |
+
post_oneshot_calibration: true
|
| 23 |
+
scheme_overrides:
|
| 24 |
+
Linear:
|
| 25 |
+
weights:
|
| 26 |
+
num_bits: 8
|
| 27 |
+
symmetric: true
|
| 28 |
+
strategy: channel
|
| 29 |
+
MatMulLeftInput_QK:
|
| 30 |
+
input_activations:
|
| 31 |
+
num_bits: 8
|
| 32 |
+
symmetric: true
|
| 33 |
+
MatMulLeftInput_PV:
|
| 34 |
+
input_activations:
|
| 35 |
+
num_bits: 8
|
| 36 |
+
symmetric: true
|
| 37 |
+
Embedding:
|
| 38 |
+
input_activations: null
|
| 39 |
+
weights:
|
| 40 |
+
num_bits: 8
|
| 41 |
+
symmetric: false
|
| 42 |
+
SparseGPTModifier:
|
| 43 |
+
sparsity: 0.0
|
| 44 |
+
block_size: 128
|
| 45 |
+
sequential_update: true
|
| 46 |
+
quantize: true
|
| 47 |
+
percdamp: 0.01
|
| 48 |
+
mask_structure: "0:0"
|
| 49 |
+
targets: ["re:model.layers.\\d*$"]
|