doppler-gpu 0.4.10 → 0.4.11

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.
Files changed (73) hide show
  1. package/CHANGELOG.md +34 -0
  2. package/README.md +27 -0
  3. package/models/catalog.json +612 -7
  4. package/package.json +13 -1
  5. package/src/client/doppler-registry.js +6 -0
  6. package/src/client/doppler-registry.json +126 -4
  7. package/src/client/runtime/lora.d.ts +10 -0
  8. package/src/client/runtime/lora.js +64 -0
  9. package/src/client/runtime/model-manager.js +6 -167
  10. package/src/client/runtime/model-session.d.ts +3 -1
  11. package/src/client/runtime/model-session.js +6 -0
  12. package/src/client/wrap-pipeline-handle.js +7 -0
  13. package/src/config/conversion/amplify/amplify-120m-f16-af32.json +516 -0
  14. package/src/config/conversion/esm/esm2-t12-35m-ur50d-f32-af32.json +571 -0
  15. package/src/config/conversion/esmc/esmc-300m-f32-af32.json +342 -0
  16. package/src/config/conversion/nucleotide-transformer/nucleotide-transformer-v2-50m-f32-af32.json +388 -0
  17. package/src/config/kernels/registry.json +52 -9
  18. package/src/config/merge.js +24 -0
  19. package/src/config/required-inference-fields-contract-check.js +2 -0
  20. package/src/config/schema/converter.schema.d.ts +7 -0
  21. package/src/config/schema/converter.schema.js +15 -0
  22. package/src/config/schema/inference.schema.d.ts +8 -0
  23. package/src/config/schema/manifest.schema.d.ts +32 -1
  24. package/src/config/schema/manifest.schema.js +26 -0
  25. package/src/converter/conversion-plan.js +15 -1
  26. package/src/converter/core.js +5 -4
  27. package/src/converter/source-tensor-rules.d.ts +50 -0
  28. package/src/converter/source-tensor-rules.js +180 -0
  29. package/src/converter/source-tokenizer.d.ts +45 -0
  30. package/src/converter/source-tokenizer.js +155 -0
  31. package/src/experimental/adapters/lora-loader.js +15 -0
  32. package/src/experimental/bridge/trainer-artifact-bridge.d.ts +2 -1
  33. package/src/experimental/bridge/trainer-artifact-bridge.js +35 -13
  34. package/src/inference/pipelines/text/attention/attn-config.js +2 -0
  35. package/src/inference/pipelines/text/attention/projections.js +161 -24
  36. package/src/inference/pipelines/text/attention/record.js +67 -9
  37. package/src/inference/pipelines/text/attention/run.js +60 -7
  38. package/src/inference/pipelines/text/attention/types.js +18 -9
  39. package/src/inference/pipelines/text/config.d.ts +5 -0
  40. package/src/inference/pipelines/text/config.js +91 -0
  41. package/src/inference/pipelines/text/ffn/dense.js +189 -2
  42. package/src/inference/pipelines/text/ffn/standard.js +50 -3
  43. package/src/inference/pipelines/text/generator-helpers.js +11 -1
  44. package/src/inference/pipelines/text/generator-runtime.js +63 -15
  45. package/src/inference/pipelines/text/generator-steps.js +3 -1
  46. package/src/inference/pipelines/text/generator.js +90 -3
  47. package/src/inference/pipelines/text/init-qkv-fusion.js +85 -1
  48. package/src/inference/pipelines/text/init.d.ts +2 -0
  49. package/src/inference/pipelines/text/init.js +2 -0
  50. package/src/inference/pipelines/text/layer.js +6 -0
  51. package/src/inference/pipelines/text/logits/cpu.d.ts +6 -0
  52. package/src/inference/pipelines/text/logits/cpu.js +22 -0
  53. package/src/inference/pipelines/text/logits/index.d.ts +1 -1
  54. package/src/inference/pipelines/text/logits/index.js +85 -11
  55. package/src/inference/pipelines/text/logits/types.d.ts +4 -0
  56. package/src/inference/pipelines/text/ops.js +32 -0
  57. package/src/inference/pipelines/text/types.d.ts +38 -0
  58. package/src/inference/pipelines/text/weights.d.ts +10 -0
  59. package/src/inference/pipelines/text/weights.js +33 -1
  60. package/src/inference/pipelines/text.d.ts +4 -2
  61. package/src/inference/pipelines/text.js +112 -0
  62. package/src/inference/tokenizers/behavior-flags.js +5 -1
  63. package/src/inference/tokenizers/bundled.js +99 -8
  64. package/src/loader/doppler-loader.d.ts +2 -0
  65. package/src/loader/doppler-loader.js +9 -0
  66. package/src/loader/final-weights-loader.d.ts +8 -0
  67. package/src/loader/final-weights-loader.js +42 -0
  68. package/src/loader/layer-loader.d.ts +6 -0
  69. package/src/loader/layer-loader.js +102 -3
  70. package/src/loader/loader-types.d.ts +15 -2
  71. package/src/tooling/node-converter.js +33 -1
  72. package/src/tooling/node-webgpu.js +85 -2
  73. package/src/version.js +1 -1
@@ -1,16 +1,28 @@
1
1
  {
2
2
  "version": 1,
3
3
  "lifecycleSchemaVersion": 1,
4
- "updatedAt": "2026-07-11",
4
+ "updatedAt": "2026-07-19",
5
5
  "models": [
6
6
  {
7
7
  "modelId": "gemma-3-270m-it-q4k-ehf16-af32",
8
8
  "family": "gemma3",
9
+ "classification": {
10
+ "domain": "language",
11
+ "tasks": [
12
+ "generation"
13
+ ],
14
+ "architectureRole": "autoregressive-decoder",
15
+ "inputs": [
16
+ "text"
17
+ ],
18
+ "outputs": [
19
+ "text"
20
+ ]
21
+ },
9
22
  "label": "Gemma 3 270M Instruct (Q4K/F32a)",
10
23
  "description": "Gemma 3 270M text model. Q4K weights, F16 embeddings/head, F32 accumulation.",
11
24
  "modes": [
12
- "text",
13
- "vision"
25
+ "text"
14
26
  ],
15
27
  "aliases": [
16
28
  "gemma3-270m",
@@ -86,11 +98,23 @@
86
98
  {
87
99
  "modelId": "gemma-3-270m-it-f16-af32",
88
100
  "family": "gemma3",
101
+ "classification": {
102
+ "domain": "language",
103
+ "tasks": [
104
+ "generation"
105
+ ],
106
+ "architectureRole": "autoregressive-decoder",
107
+ "inputs": [
108
+ "text"
109
+ ],
110
+ "outputs": [
111
+ "text"
112
+ ]
113
+ },
89
114
  "label": "Gemma 3 270M Instruct (F16/F32a)",
90
115
  "description": "Gemma 3 270M text model. F16 weights, F16 embeddings/head, F32 accumulation.",
91
116
  "modes": [
92
- "text",
93
- "vision"
117
+ "text"
94
118
  ],
95
119
  "aliases": [
96
120
  "gemma-3-270m-it-f16",
@@ -146,6 +170,19 @@
146
170
  {
147
171
  "modelId": "google-embeddinggemma-300m-q4k-ehf16-af32",
148
172
  "family": "embeddinggemma",
173
+ "classification": {
174
+ "domain": "language",
175
+ "tasks": [
176
+ "embedding"
177
+ ],
178
+ "architectureRole": "decoder-embedding-encoder",
179
+ "inputs": [
180
+ "text"
181
+ ],
182
+ "outputs": [
183
+ "pooled-embedding"
184
+ ]
185
+ },
149
186
  "label": "EmbeddingGemma 300M (Q4K)",
150
187
  "description": "EmbeddingGemma 300M sentence embedding model. Q4K weights, F16 embeddings/head, F32 accumulation.",
151
188
  "modes": [
@@ -211,11 +248,23 @@
211
248
  {
212
249
  "modelId": "gemma-3-1b-it-q4k-ehf16-af32",
213
250
  "family": "gemma3",
251
+ "classification": {
252
+ "domain": "language",
253
+ "tasks": [
254
+ "generation"
255
+ ],
256
+ "architectureRole": "autoregressive-decoder",
257
+ "inputs": [
258
+ "text"
259
+ ],
260
+ "outputs": [
261
+ "text"
262
+ ]
263
+ },
214
264
  "label": "Gemma 3 1B Instruct (Q4K/F32a)",
215
265
  "description": "Gemma 3 1B text model. Q4K weights, F16 embeddings/head, F32 accumulation.",
216
266
  "modes": [
217
- "text",
218
- "vision"
267
+ "text"
219
268
  ],
220
269
  "aliases": [
221
270
  "gemma3-1b",
@@ -279,6 +328,21 @@
279
328
  {
280
329
  "modelId": "gemma-4-e2b-it-q4k-ehf16-af32",
281
330
  "family": "gemma4",
331
+ "classification": {
332
+ "domain": "language",
333
+ "tasks": [
334
+ "generation"
335
+ ],
336
+ "architectureRole": "autoregressive-decoder",
337
+ "inputs": [
338
+ "text",
339
+ "image",
340
+ "audio"
341
+ ],
342
+ "outputs": [
343
+ "text"
344
+ ]
345
+ },
282
346
  "label": "Gemma 4 E2B Instruct (Q4K/F32a)",
283
347
  "description": "Gemma 4 E2B multimodal model. Q4K weights, F16 embeddings/head/vision/audio/projector, F32 accumulation.",
284
348
  "modes": [
@@ -345,6 +409,21 @@
345
409
  {
346
410
  "modelId": "gemma-4-e2b-it-q4k-ehf16-af32-int4ple",
347
411
  "family": "gemma4",
412
+ "classification": {
413
+ "domain": "language",
414
+ "tasks": [
415
+ "generation"
416
+ ],
417
+ "architectureRole": "autoregressive-decoder",
418
+ "inputs": [
419
+ "text",
420
+ "image",
421
+ "audio"
422
+ ],
423
+ "outputs": [
424
+ "text"
425
+ ]
426
+ },
348
427
  "label": "Gemma 4 E2B Instruct (Q4K/F32a/INT4 PLE)",
349
428
  "description": "Gemma 4 E2B multimodal model. Q4K weights, INT4 per-layer embeddings with per-row F32 scales, F16 embeddings/head/vision/audio/projector, F32 accumulation.",
350
429
  "modes": [
@@ -415,6 +494,21 @@
415
494
  {
416
495
  "modelId": "gemma-4-e2b-it-q4k-ehf16-af16-int4ple",
417
496
  "family": "gemma4",
497
+ "classification": {
498
+ "domain": "language",
499
+ "tasks": [
500
+ "generation"
501
+ ],
502
+ "architectureRole": "autoregressive-decoder",
503
+ "inputs": [
504
+ "text",
505
+ "image",
506
+ "audio"
507
+ ],
508
+ "outputs": [
509
+ "text"
510
+ ]
511
+ },
418
512
  "label": "Gemma 4 E2B Instruct (Q4K/F16a/INT4 PLE)",
419
513
  "description": "Gemma 4 E2B multimodal Doppler/WebGPU f16 activation variant. Shares the existing Q4K/INT4 PLE weight pack and falls back to AF32 when capability rules reject AF16.",
420
514
  "modes": [
@@ -482,6 +576,19 @@
482
576
  {
483
577
  "modelId": "gemma-4-12b-it-text-q4k-ehf16-af32",
484
578
  "family": "gemma4",
579
+ "classification": {
580
+ "domain": "language",
581
+ "tasks": [
582
+ "generation"
583
+ ],
584
+ "architectureRole": "autoregressive-decoder",
585
+ "inputs": [
586
+ "text"
587
+ ],
588
+ "outputs": [
589
+ "text"
590
+ ]
591
+ },
485
592
  "label": "Gemma 4 12B Instruct Text (Q4K/F32a)",
486
593
  "description": "Gemma 4 12B text-only local RDRR lane. Q4K weights, F16 embeddings/head, F32 activation contract.",
487
594
  "modes": [
@@ -542,6 +649,19 @@
542
649
  {
543
650
  "modelId": "gemma-4-12b-it-text-q4k-ehf16-af16",
544
651
  "family": "gemma4",
652
+ "classification": {
653
+ "domain": "language",
654
+ "tasks": [
655
+ "generation"
656
+ ],
657
+ "architectureRole": "autoregressive-decoder",
658
+ "inputs": [
659
+ "text"
660
+ ],
661
+ "outputs": [
662
+ "text"
663
+ ]
664
+ },
545
665
  "label": "Gemma 4 12B Instruct Text (Q4K/F16a)",
546
666
  "description": "Gemma 4 12B text-only local f16 activation sibling over the existing Q4K weight pack.",
547
667
  "modes": [
@@ -602,6 +722,19 @@
602
722
  {
603
723
  "modelId": "gemma-4-12b-it-text-w4a16-ct-ehf16-af16",
604
724
  "family": "gemma4",
725
+ "classification": {
726
+ "domain": "language",
727
+ "tasks": [
728
+ "generation"
729
+ ],
730
+ "architectureRole": "autoregressive-decoder",
731
+ "inputs": [
732
+ "text"
733
+ ],
734
+ "outputs": [
735
+ "text"
736
+ ]
737
+ },
605
738
  "label": "Gemma 4 12B Instruct Text (W4A16 QAT/F16a)",
606
739
  "description": "Gemma 4 12B text-only RDRR lane from Google's W4A16 QAT compressed-tensors checkpoint. W4A16 weights, F16 embeddings/head, F16 activation contract.",
607
740
  "modes": [
@@ -663,6 +796,19 @@
663
796
  {
664
797
  "modelId": "diffusiongemma-26b-a4b-it-q4k-ehf16-af16",
665
798
  "family": "diffusiongemma",
799
+ "classification": {
800
+ "domain": "language",
801
+ "tasks": [
802
+ "generation"
803
+ ],
804
+ "architectureRole": "block-diffusion-decoder",
805
+ "inputs": [
806
+ "text"
807
+ ],
808
+ "outputs": [
809
+ "text"
810
+ ]
811
+ },
666
812
  "label": "DiffusionGemma 26B A4B Instruct (Q4K/F16a)",
667
813
  "description": "DiffusionGemma 26B A4B text-only block-diffusion MoE artifact. Q4K weights, F16 embeddings/head, F16 activation contract, and DiffusionGemma throughput runtime profile.",
668
814
  "modes": [
@@ -709,6 +855,19 @@
709
855
  {
710
856
  "modelId": "gemma-4-31b-it-text-q4k-ehf16-af32",
711
857
  "family": "gemma4",
858
+ "classification": {
859
+ "domain": "language",
860
+ "tasks": [
861
+ "generation"
862
+ ],
863
+ "architectureRole": "autoregressive-decoder",
864
+ "inputs": [
865
+ "text"
866
+ ],
867
+ "outputs": [
868
+ "text"
869
+ ]
870
+ },
712
871
  "label": "Gemma 4 31B Instruct Text (Q4K/F32a)",
713
872
  "description": "Gemma 4 31B text-only model. Q4K weights, F16 embeddings/head, F32 accumulation.",
714
873
  "modes": [
@@ -774,6 +933,19 @@
774
933
  {
775
934
  "modelId": "gemma-4-31b-it-text-q4k-ehf16-af16",
776
935
  "family": "gemma4",
936
+ "classification": {
937
+ "domain": "language",
938
+ "tasks": [
939
+ "generation"
940
+ ],
941
+ "architectureRole": "autoregressive-decoder",
942
+ "inputs": [
943
+ "text"
944
+ ],
945
+ "outputs": [
946
+ "text"
947
+ ]
948
+ },
777
949
  "label": "Gemma 4 31B Instruct Text (Q4K/F16a)",
778
950
  "description": "Gemma 4 31B text-only Doppler/WebGPU f16 activation variant. Shares the existing Q4K/F16 embedding and head weight pack; Doe/Cerebras f16 CSL lowering is intentionally unsupported.",
779
951
  "modes": [
@@ -834,6 +1006,19 @@
834
1006
  {
835
1007
  "modelId": "translategemma-4b-1b-enes-q4k-ehf16-af32",
836
1008
  "family": "translategemma",
1009
+ "classification": {
1010
+ "domain": "language",
1011
+ "tasks": [
1012
+ "translation"
1013
+ ],
1014
+ "architectureRole": "autoregressive-decoder",
1015
+ "inputs": [
1016
+ "text"
1017
+ ],
1018
+ "outputs": [
1019
+ "text"
1020
+ ]
1021
+ },
837
1022
  "label": "TranslateGemma 1B EN/ES Student (NativeKD2 Q4K)",
838
1023
  "description": "Experimental-tier English/Spanish translation specialist distilled from TranslateGemma 4B into Gemma 3 1B. The hosted Q4K artifact has deterministic WMT13 and browser/WebGPU evidence; quality and performance claims remain limited to the recorded EN/ES and Radeon/Vulkan receipts.",
839
1024
  "modes": [
@@ -898,6 +1083,19 @@
898
1083
  {
899
1084
  "modelId": "translategemma-4b-it-q4k-ehf16-af32",
900
1085
  "family": "translategemma",
1086
+ "classification": {
1087
+ "domain": "language",
1088
+ "tasks": [
1089
+ "translation"
1090
+ ],
1091
+ "architectureRole": "autoregressive-decoder",
1092
+ "inputs": [
1093
+ "text"
1094
+ ],
1095
+ "outputs": [
1096
+ "text"
1097
+ ]
1098
+ },
901
1099
  "label": "TranslateGemma 4B Instruct (Q4K)",
902
1100
  "description": "TranslateGemma 4B translation model. Q4K weights, F16 embeddings/head, F32 accumulation.",
903
1101
  "modes": [
@@ -964,6 +1162,19 @@
964
1162
  {
965
1163
  "modelId": "qwen-3-embedding-0-6b-q4k-ehf16-af32",
966
1164
  "family": "qwen3",
1165
+ "classification": {
1166
+ "domain": "language",
1167
+ "tasks": [
1168
+ "embedding"
1169
+ ],
1170
+ "architectureRole": "decoder-embedding-encoder",
1171
+ "inputs": [
1172
+ "text"
1173
+ ],
1174
+ "outputs": [
1175
+ "pooled-embedding"
1176
+ ]
1177
+ },
967
1178
  "label": "Qwen3 Embedding 0.6B (Q4K/F32a)",
968
1179
  "description": "Qwen3 0.6B SentenceTransformers embedding target with manifest-owned last-token pooling, prompt inclusion, and L2 normalization. Node/WebGPU verification passes semantic retrieval and pair thresholds; the local compare lane cites the Transformers.js ONNX WebGPU baseline.",
969
1180
  "modes": [
@@ -1039,6 +1250,19 @@
1039
1250
  {
1040
1251
  "modelId": "qwen-3-reranker-0-6b-q4k-ehf16-af32",
1041
1252
  "family": "qwen3",
1253
+ "classification": {
1254
+ "domain": "language",
1255
+ "tasks": [
1256
+ "reranking"
1257
+ ],
1258
+ "architectureRole": "decoder-cross-encoder",
1259
+ "inputs": [
1260
+ "text-pair"
1261
+ ],
1262
+ "outputs": [
1263
+ "relevance-score"
1264
+ ]
1265
+ },
1042
1266
  "label": "Qwen3 Reranker 0.6B (Q4K/F32a)",
1043
1267
  "description": "Qwen3 0.6B SentenceTransformers CrossEncoder reranker target with Q4K attention, F16 MLP projections, and manifest-owned true-logit scoring. The hosted registry lane passes semantic rerank verification and has an external Transformers.js ONNX compare profile for fair rerank receipts.",
1044
1268
  "modes": [
@@ -1123,6 +1347,19 @@
1123
1347
  {
1124
1348
  "modelId": "qwen-3-reranker-0-6b-f16-af32",
1125
1349
  "family": "qwen3",
1350
+ "classification": {
1351
+ "domain": "language",
1352
+ "tasks": [
1353
+ "reranking"
1354
+ ],
1355
+ "architectureRole": "decoder-cross-encoder",
1356
+ "inputs": [
1357
+ "text-pair"
1358
+ ],
1359
+ "outputs": [
1360
+ "relevance-score"
1361
+ ]
1362
+ },
1126
1363
  "label": "Qwen3 Reranker 0.6B (F16/F32a)",
1127
1364
  "description": "Qwen3 0.6B SentenceTransformers CrossEncoder reranker target with manifest-owned true/false token scoring. The F16 lane passes the semantic rerank suite and is the supported reranker variant for publication.",
1128
1365
  "modes": [
@@ -1200,6 +1437,19 @@
1200
1437
  {
1201
1438
  "modelId": "minicpm4-0-5b-f16-af32",
1202
1439
  "family": "minicpm",
1440
+ "classification": {
1441
+ "domain": "language",
1442
+ "tasks": [
1443
+ "generation"
1444
+ ],
1445
+ "architectureRole": "autoregressive-decoder",
1446
+ "inputs": [
1447
+ "text"
1448
+ ],
1449
+ "outputs": [
1450
+ "text"
1451
+ ]
1452
+ },
1203
1453
  "label": "MiniCPM4 0.5B (F16/F32a)",
1204
1454
  "description": "MiniCPM4 0.5B text onboarding artifact. The manifest and shards are complete, but the deterministic smoke prompt produced incoherent text, so this lane is not a supported generation model yet.",
1205
1455
  "modes": [
@@ -1257,6 +1507,19 @@
1257
1507
  {
1258
1508
  "modelId": "qwen-3-5-0-8b-q4k-ehaf16",
1259
1509
  "family": "qwen3",
1510
+ "classification": {
1511
+ "domain": "language",
1512
+ "tasks": [
1513
+ "generation"
1514
+ ],
1515
+ "architectureRole": "autoregressive-decoder",
1516
+ "inputs": [
1517
+ "text"
1518
+ ],
1519
+ "outputs": [
1520
+ "text"
1521
+ ]
1522
+ },
1260
1523
  "label": "Qwen 3.5 0.8B (Q4K)",
1261
1524
  "description": "Qwen 3.5 0.8B text model with hybrid linear attention. Q4K weights, F16 embeddings, Q4 tied LM head, F32 accumulation.",
1262
1525
  "modes": [
@@ -1346,6 +1609,19 @@
1346
1609
  {
1347
1610
  "modelId": "qwen-3-5-2b-q4k-ehaf16",
1348
1611
  "family": "qwen3",
1612
+ "classification": {
1613
+ "domain": "language",
1614
+ "tasks": [
1615
+ "generation"
1616
+ ],
1617
+ "architectureRole": "autoregressive-decoder",
1618
+ "inputs": [
1619
+ "text"
1620
+ ],
1621
+ "outputs": [
1622
+ "text"
1623
+ ]
1624
+ },
1349
1625
  "label": "Qwen 3.5 2B (Q4K)",
1350
1626
  "description": "Qwen 3.5 2B text model with hybrid linear attention. Q4K weights, F16 embeddings, Q4 tied LM head, F32 accumulation.",
1351
1627
  "modes": [
@@ -1417,6 +1693,19 @@
1417
1693
  {
1418
1694
  "modelId": "qwen-3-6-27b-q4k-ehaf16",
1419
1695
  "family": "qwen3",
1696
+ "classification": {
1697
+ "domain": "language",
1698
+ "tasks": [
1699
+ "generation"
1700
+ ],
1701
+ "architectureRole": "autoregressive-decoder",
1702
+ "inputs": [
1703
+ "text"
1704
+ ],
1705
+ "outputs": [
1706
+ "text"
1707
+ ]
1708
+ },
1420
1709
  "label": "Qwen 3.6 27B (Q4K)",
1421
1710
  "description": "Qwen 3.6 27B text model with hybrid linear attention. Q4K weights, F16 embeddings, F32 accumulation.",
1422
1711
  "modes": [
@@ -1483,6 +1772,19 @@
1483
1772
  {
1484
1773
  "modelId": "qwen-3-6-27b-q4k-eaf16",
1485
1774
  "family": "qwen3",
1775
+ "classification": {
1776
+ "domain": "language",
1777
+ "tasks": [
1778
+ "generation"
1779
+ ],
1780
+ "architectureRole": "autoregressive-decoder",
1781
+ "inputs": [
1782
+ "text"
1783
+ ],
1784
+ "outputs": [
1785
+ "text"
1786
+ ]
1787
+ },
1486
1788
  "label": "Qwen 3.6 27B (Q4K/F16a)",
1487
1789
  "description": "Qwen 3.6 27B text-only Doppler/WebGPU f16 lane. Shares the existing Q4K/F16 embedding weight pack; Doe/Cerebras f16 CSL lowering is intentionally unsupported.",
1488
1790
  "modes": [
@@ -1544,6 +1846,19 @@
1544
1846
  {
1545
1847
  "modelId": "lfm2-5-1-2b-instruct-q4k-ehf16-af32",
1546
1848
  "family": "lfm2",
1849
+ "classification": {
1850
+ "domain": "language",
1851
+ "tasks": [
1852
+ "generation"
1853
+ ],
1854
+ "architectureRole": "autoregressive-decoder",
1855
+ "inputs": [
1856
+ "text"
1857
+ ],
1858
+ "outputs": [
1859
+ "text"
1860
+ ]
1861
+ },
1547
1862
  "label": "LFM2.5 1.2B Instruct (Q4K/F32a)",
1548
1863
  "description": "Liquid AI LFM2.5 1.2B Instruct text model. Q4K weights, F16 embeddings/head, F32 accumulation.",
1549
1864
  "modes": [
@@ -1584,6 +1899,296 @@
1584
1899
  "artifactCompleteness": "incomplete",
1585
1900
  "runtimePromotionState": "unpromoted",
1586
1901
  "weightsRefAllowed": false
1902
+ },
1903
+ {
1904
+ "modelId": "amplify-120m-f16-af32",
1905
+ "family": "amplify",
1906
+ "classification": {
1907
+ "domain": "protein",
1908
+ "tasks": [
1909
+ "embedding",
1910
+ "masked-token-prediction"
1911
+ ],
1912
+ "architectureRole": "bidirectional-encoder",
1913
+ "inputs": [
1914
+ "protein-sequence"
1915
+ ],
1916
+ "outputs": [
1917
+ "token-embedding",
1918
+ "pooled-embedding",
1919
+ "token-logits"
1920
+ ]
1921
+ },
1922
+ "label": "AMPLIFY 120M Protein Encoder (F16/F32a)",
1923
+ "description": "AMPLIFY 120M protein-sequence encoder with F16 weights and F32 activations.",
1924
+ "modes": [
1925
+ "embedding"
1926
+ ],
1927
+ "aliases": [
1928
+ "amplify-120m",
1929
+ "chandar-lab/AMPLIFY_120M"
1930
+ ],
1931
+ "artifact": {
1932
+ "format": "rdrr"
1933
+ },
1934
+ "hf": {
1935
+ "repoId": "Clocksmith/rdrr",
1936
+ "path": "models/amplify-120m-f16-af32",
1937
+ "revision": "990de206946a0972cc92e9dbcc9726399d257cac"
1938
+ },
1939
+ "baseUrl": null,
1940
+ "quickstart": false,
1941
+ "sizeBytes": 236647491,
1942
+ "recommended": false,
1943
+ "sortOrder": 36,
1944
+ "lifecycle": {
1945
+ "availability": {
1946
+ "curated": false,
1947
+ "local": false,
1948
+ "hf": true
1949
+ },
1950
+ "status": {
1951
+ "runtime": "active",
1952
+ "conversion": "ready",
1953
+ "demo": "none",
1954
+ "tested": "verified"
1955
+ },
1956
+ "tested": {
1957
+ "suite": "sequence-embedding",
1958
+ "surface": [
1959
+ "node"
1960
+ ],
1961
+ "result": "pass",
1962
+ "lastVerifiedAt": "2026-07-19",
1963
+ "source": "runtime-verify",
1964
+ "notes": "Hosted Node/WebGPU sequence parity and synthetic q_proj LoRA lifecycle passed at docs/status/amplify-120m-sequence-hosted-webgpu-lora-qualification-2026-07-19.json. This is not a downstream biological-task quality claim.",
1965
+ "contracts": {
1966
+ "executionContractOk": true
1967
+ }
1968
+ }
1969
+ },
1970
+ "sourceCheckpointId": "chandar-lab/AMPLIFY_120M@d918a9e8c64dc43f6ec4996022c3deba14d66470",
1971
+ "weightPackId": "amplify-120m-f16-af32-wp-v1",
1972
+ "manifestVariantId": "amplify-120m-f16-af32-mv-exec-v1",
1973
+ "artifactCompleteness": "complete",
1974
+ "runtimePromotionState": "manifest-owned",
1975
+ "weightsRefAllowed": false
1976
+ },
1977
+ {
1978
+ "modelId": "esm2-t12-35m-ur50d-f32-af32",
1979
+ "family": "esm",
1980
+ "classification": {
1981
+ "domain": "protein",
1982
+ "tasks": [
1983
+ "embedding"
1984
+ ],
1985
+ "architectureRole": "bidirectional-encoder",
1986
+ "inputs": [
1987
+ "protein-sequence"
1988
+ ],
1989
+ "outputs": [
1990
+ "token-embedding",
1991
+ "pooled-embedding"
1992
+ ]
1993
+ },
1994
+ "label": "ESM-2 35M Protein Encoder (F32)",
1995
+ "description": "ESM-2 35M protein-sequence encoder with F32 weights and activations.",
1996
+ "modes": [
1997
+ "embedding"
1998
+ ],
1999
+ "aliases": [
2000
+ "esm2-35m",
2001
+ "facebook/esm2_t12_35M_UR50D"
2002
+ ],
2003
+ "artifact": {
2004
+ "format": "rdrr"
2005
+ },
2006
+ "hf": {
2007
+ "repoId": "Clocksmith/rdrr",
2008
+ "path": "models/esm2-t12-35m-ur50d-f32-af32",
2009
+ "revision": "feea324bae6c06c9d129ce39e188a75478ccdbbc"
2010
+ },
2011
+ "baseUrl": null,
2012
+ "quickstart": false,
2013
+ "sizeBytes": 133159807,
2014
+ "recommended": false,
2015
+ "sortOrder": 37,
2016
+ "lifecycle": {
2017
+ "availability": {
2018
+ "curated": false,
2019
+ "local": false,
2020
+ "hf": true
2021
+ },
2022
+ "status": {
2023
+ "runtime": "active",
2024
+ "conversion": "ready",
2025
+ "demo": "none",
2026
+ "tested": "verified"
2027
+ },
2028
+ "tested": {
2029
+ "suite": "sequence-embedding",
2030
+ "surface": [
2031
+ "node"
2032
+ ],
2033
+ "result": "pass",
2034
+ "lastVerifiedAt": "2026-07-19",
2035
+ "source": "runtime-verify",
2036
+ "notes": "Hosted Node/WebGPU sequence parity and synthetic q_proj LoRA lifecycle passed at docs/status/esm2-t12-35m-sequence-hosted-webgpu-lora-qualification-2026-07-19.json. The initial lane exposes embeddings, not masked-token logits or downstream biological quality.",
2037
+ "contracts": {
2038
+ "executionContractOk": true
2039
+ }
2040
+ }
2041
+ },
2042
+ "sourceCheckpointId": "facebook/esm2_t12_35M_UR50D@6fbf070e65b0b7291e7bbcd451118c216cff79d8",
2043
+ "weightPackId": "esm2-t12-35m-ur50d-f32-af32-wp-v1",
2044
+ "manifestVariantId": "esm2-t12-35m-ur50d-f32-af32-mv-exec-v1",
2045
+ "artifactCompleteness": "complete",
2046
+ "runtimePromotionState": "manifest-owned",
2047
+ "weightsRefAllowed": false
2048
+ },
2049
+ {
2050
+ "modelId": "esmc-300m-f32-af32",
2051
+ "family": "esmc",
2052
+ "classification": {
2053
+ "domain": "protein",
2054
+ "tasks": [
2055
+ "embedding"
2056
+ ],
2057
+ "architectureRole": "bidirectional-encoder",
2058
+ "inputs": [
2059
+ "protein-sequence"
2060
+ ],
2061
+ "outputs": [
2062
+ "token-embedding",
2063
+ "pooled-embedding"
2064
+ ]
2065
+ },
2066
+ "label": "ESMC 300M Protein Encoder (F32)",
2067
+ "description": "ESMC 300M protein-sequence encoder with F32 weights and activations.",
2068
+ "modes": [
2069
+ "embedding"
2070
+ ],
2071
+ "aliases": [
2072
+ "esmc-300m",
2073
+ "biohub/ESMC-300M"
2074
+ ],
2075
+ "artifact": {
2076
+ "format": "rdrr"
2077
+ },
2078
+ "hf": {
2079
+ "repoId": "Clocksmith/rdrr",
2080
+ "path": "models/esmc-300m-f32-af32",
2081
+ "revision": "8354795dc544798ca7a4419bf00f844648f4b6f0"
2082
+ },
2083
+ "baseUrl": null,
2084
+ "quickstart": false,
2085
+ "sizeBytes": 1328207519,
2086
+ "recommended": false,
2087
+ "sortOrder": 38,
2088
+ "lifecycle": {
2089
+ "availability": {
2090
+ "curated": false,
2091
+ "local": false,
2092
+ "hf": true
2093
+ },
2094
+ "status": {
2095
+ "runtime": "active",
2096
+ "conversion": "ready",
2097
+ "demo": "none",
2098
+ "tested": "verified"
2099
+ },
2100
+ "tested": {
2101
+ "suite": "sequence-embedding",
2102
+ "surface": [
2103
+ "node"
2104
+ ],
2105
+ "result": "pass",
2106
+ "lastVerifiedAt": "2026-07-19",
2107
+ "source": "runtime-verify",
2108
+ "notes": "Hosted Node/WebGPU sequence parity and synthetic q_proj LoRA lifecycle passed at docs/status/esmc-300m-sequence-hosted-webgpu-lora-qualification-2026-07-19.json. The initial lane exposes embeddings, not masked-token logits or downstream biological quality.",
2109
+ "contracts": {
2110
+ "executionContractOk": true
2111
+ }
2112
+ }
2113
+ },
2114
+ "sourceCheckpointId": "biohub/ESMC-300M@a59b831785f907e96e6a246b1d142bfb76df31ee",
2115
+ "weightPackId": "esmc-300m-f32-af32-wp-v1",
2116
+ "manifestVariantId": "esmc-300m-f32-af32-mv-exec-v1",
2117
+ "artifactCompleteness": "complete",
2118
+ "runtimePromotionState": "manifest-owned",
2119
+ "weightsRefAllowed": false
2120
+ },
2121
+ {
2122
+ "modelId": "nucleotide-transformer-v2-50m-f32-af32",
2123
+ "family": "nucleotide-transformer",
2124
+ "classification": {
2125
+ "domain": "nucleotide",
2126
+ "tasks": [
2127
+ "embedding"
2128
+ ],
2129
+ "architectureRole": "bidirectional-encoder",
2130
+ "inputs": [
2131
+ "dna-sequence"
2132
+ ],
2133
+ "outputs": [
2134
+ "token-embedding",
2135
+ "pooled-embedding"
2136
+ ]
2137
+ },
2138
+ "label": "Nucleotide Transformer v2 50M DNA Encoder (F32)",
2139
+ "description": "Nucleotide Transformer v2 50M DNA-sequence encoder with F32 weights and activations.",
2140
+ "modes": [
2141
+ "embedding"
2142
+ ],
2143
+ "aliases": [
2144
+ "nucleotide-transformer-v2-50m",
2145
+ "InstaDeepAI/nucleotide-transformer-v2-50m-multi-species"
2146
+ ],
2147
+ "artifact": {
2148
+ "format": "rdrr"
2149
+ },
2150
+ "hf": {
2151
+ "repoId": "Clocksmith/rdrr",
2152
+ "path": "models/nucleotide-transformer-v2-50m-f32-af32",
2153
+ "revision": "a5703ef6971cc883438a393d73f2f54dccf94976"
2154
+ },
2155
+ "baseUrl": null,
2156
+ "quickstart": false,
2157
+ "sizeBytes": 210074179,
2158
+ "recommended": false,
2159
+ "sortOrder": 39,
2160
+ "lifecycle": {
2161
+ "availability": {
2162
+ "curated": false,
2163
+ "local": false,
2164
+ "hf": true
2165
+ },
2166
+ "status": {
2167
+ "runtime": "active",
2168
+ "conversion": "ready",
2169
+ "demo": "none",
2170
+ "tested": "verified"
2171
+ },
2172
+ "tested": {
2173
+ "suite": "sequence-embedding",
2174
+ "surface": [
2175
+ "node"
2176
+ ],
2177
+ "result": "pass",
2178
+ "lastVerifiedAt": "2026-07-19",
2179
+ "source": "runtime-verify",
2180
+ "notes": "Hosted Node/WebGPU sequence parity and synthetic q_proj LoRA lifecycle passed at docs/status/nucleotide-transformer-v2-50m-sequence-hosted-webgpu-lora-qualification-2026-07-19.json. The initial lane exposes embeddings, not masked-token logits or downstream biological quality; upstream use is CC-BY-NC-SA-4.0.",
2181
+ "contracts": {
2182
+ "executionContractOk": true
2183
+ }
2184
+ }
2185
+ },
2186
+ "sourceCheckpointId": "InstaDeepAI/nucleotide-transformer-v2-50m-multi-species@81b29e5786726d891dbf929404ef20adca5b36f1",
2187
+ "weightPackId": "nucleotide-transformer-v2-50m-f32-af32-wp-v1",
2188
+ "manifestVariantId": "nucleotide-transformer-v2-50m-f32-af32-mv-exec-v1",
2189
+ "artifactCompleteness": "complete",
2190
+ "runtimePromotionState": "manifest-owned",
2191
+ "weightsRefAllowed": false
1587
2192
  }
1588
2193
  ]
1589
2194
  }