@daocloud-proto/hydra 0.0.0-dev-29dd9c89 → 0.0.0-dev-c571dd90

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -30,6 +30,44 @@ export enum FineTuningType {
30
30
  FREEZE = "FREEZE",
31
31
  }
32
32
 
33
+ export enum ComputeType {
34
+ COMPUTE_TYPE_UNSPECIFIED = "COMPUTE_TYPE_UNSPECIFIED",
35
+ BF16 = "BF16",
36
+ FP16 = "FP16",
37
+ FP32 = "FP32",
38
+ PURE_BF16 = "PURE_BF16",
39
+ }
40
+
41
+ export enum LrSchedulerType {
42
+ LR_SCHEDULER_TYPE_UNSPECIFIED = "LR_SCHEDULER_TYPE_UNSPECIFIED",
43
+ COSINE = "COSINE",
44
+ LINEAR = "LINEAR",
45
+ COSINE_WITH_RESTARTS = "COSINE_WITH_RESTARTS",
46
+ POLYNOMIAL = "POLYNOMIAL",
47
+ CONSTANT = "CONSTANT",
48
+ CONSTANT_WITH_WARMUP = "CONSTANT_WITH_WARMUP",
49
+ INVERSE_SQRT = "INVERSE_SQRT",
50
+ REDUCE_LR_ON_PLATEAU = "REDUCE_LR_ON_PLATEAU",
51
+ COSINE_WITH_MIN_LR = "COSINE_WITH_MIN_LR",
52
+ WARMUP_STABLE_DECAY = "WARMUP_STABLE_DECAY",
53
+ }
54
+
55
+ export enum OptimizerType {
56
+ OPTIMIZER_TYPE_UNSPECIFIED = "OPTIMIZER_TYPE_UNSPECIFIED",
57
+ ADAMW_TORCH = "ADAMW_TORCH",
58
+ ADAMW_HF = "ADAMW_HF",
59
+ SGD = "SGD",
60
+ ADAFACTOR = "ADAFACTOR",
61
+ ADAGRAD = "ADAGRAD",
62
+ }
63
+
64
+ export enum QuantizationMethodType {
65
+ QUANTIZATION_METHOD_TYPE_UNSPECIFIED = "QUANTIZATION_METHOD_TYPE_UNSPECIFIED",
66
+ BITSANDBYTES = "BITSANDBYTES",
67
+ HQQ = "HQQ",
68
+ EETQ = "EETQ",
69
+ }
70
+
33
71
  export type ModelConfig = {
34
72
  modelId?: string
35
73
  modelVolumeId?: string
@@ -52,7 +90,7 @@ export type LoraConfig = {
52
90
  pissaConvert?: boolean
53
91
  createNewAdapter?: boolean
54
92
  quantizationBit?: number
55
- quantizationMethod?: string
93
+ quantizationMethod?: QuantizationMethodType
56
94
  ropeScaling?: string
57
95
  booster?: string
58
96
  }
@@ -70,11 +108,11 @@ export type TrainingConfig = {
70
108
  perDeviceTrainBatchSize?: number
71
109
  gradientAccumulationSteps?: number
72
110
  learningRate?: string
73
- lrSchedulerType?: string
111
+ lrSchedulerType?: LrSchedulerType
74
112
  warmupRatio?: string
75
113
  warmupSteps?: number
76
- bf16?: boolean
77
- optimizer?: string
114
+ computeType?: ComputeType
115
+ optimizer?: OptimizerType
78
116
  seed?: number
79
117
  ddpTimeout?: number
80
118
  maxGradNorm?: string
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@daocloud-proto/hydra",
3
- "version": "v0.0.0-dev-29dd9c89",
3
+ "version": "v0.0.0-dev-c571dd90",
4
4
  "description": "",
5
5
  "author": "",
6
6
  "license": "ISC"