## NVIDIA-NeMo/Automodel — v0.3.0rc4…v0.3.0

_51 commits._

### Features
- **cp: `ci: Add-credentials-for-docs (1389)` into `r0.3.0` (#1396)** (6bb891c)
- **cp: `ci: Add open-clip-torch as vlm dependency (1329)` into `r0.3.0` (#1331)** (27e290e)
- **cp: `feat: add transformers omni and port over qwen3omni (1321)` into `r0.3.0` (#1326)** (350276e)
- **cp: `feat: Add TE fp8 for models that support it (1205)` into `r0.3.0` (#1235)** (b43c0ea)

### Fixes
- **cp: `fix: MoE lora adapter layout (1395)` into `r0.3.0` (#1399)** (9e9472f)
- **cp: `fix: guard AutoConfig.from_pretrained in PP mask precomputation (1378)` into `r0.3.0` (#1381)** (e30fec3)
- **cp: `fix: loss masking with pad eos collision (1338)` into `r0.3.0` (#1368)** (f89c8e9)
- **cp: `fix: prefer moe_config for num_experts in apply_ac (1361)` into `r0.3.0` (#1362)** (1bd2bc0)
- **cp: `fix: FSDP pre-shard combined projections on dim 1 for Qwen2.5-7B support (1357)` into `r0.3.0` (#1363)** (a172815)
- **cp: `fix: combined projection (1324)` into `r0.3.0` (#1353)** (a6535c0)
- **cp: `feat: qol setup fix (1350)` into `r0.3.0` (#1352)** (3fc5743)
- **cp: `fix: update kimi K2.5 to transformers v5 (1337)` into `r0.3.0` (#1346)** (26814f5)
- **cp: `fix: correctly resume moe lora ckpt (1325)` into `r0.3.0` (#1341)** (821ae2a)
- **cp: fix phi4 seedoss tokenzer issues (#1267) (#1330)** (9f250ae)
- **cp: `fix: ensure image rgb (1319)` into `r0.3.0` (#1327)** (bc8bfa7)
- **fix: moonlight and qwen moe lora configs (#1328)** (4f750de)
- **cp: fix: skip load-before-shard when expert parallelism is active (#1301) (#1316)** (e2be886)
- **cp: `fix: moe lora ckpt saving and loading (1278)` into `r0.3.0` (#1312)** (30fd164)
- **cp: `fix: nightly CI fixes (1271)` into `r0.3.0` (#1302)** (9186cc3)
- **cp: `fix: do not load full sd/still use dcp for custom model (1309)` into `r0.3.0` (#1311)** (434758d)
- **cp: `fix: fix optimized tp plan lookup (1282)` into `r0.3.0` (#1284)** (5727fd2)
- **cp: `fix: reenable activation checkpointing in moe configs (1285)` into `r0.3.0` (#1289)** (53c5984)
- **cp: `fix: remove unused argument from dsv32 config (1281)` into `r0.3.0` (#1283)** (ef798b1)
- **cp: `fix: bump up liger-kernel to its transformers-5.0.0rc1 release (1212)` into `r0.3.0` (#1277)** (199d15b)
- **cp: `fix: apply param freeze the right place for moe lora  (1252)` into `r0.3.0` (#1272)** (13d2bbb)
- **cp: `fix: vlm refac fixes (1268)` into `r0.3.0` (#1273)** (320ebee)
- **cp: fix: add missing hf checkpoint mixin to custom models (#1256) (#1262)** (df68a08)
- **cp: `fix: model type detection (1251)` into `r0.3.0` (#1254)** (63db36a)
- **cp: `fix: moonlight tokenization with transformers v5 (1243)` into `r0.3.0` (#1249)** (918ef88)
- **cp: `fix: add missint tokenizer for qwen3 moe 2layer proxy torch sdpa (1247)` into `r0.3.0` (#1248)** (10a94a7)
- **cp: `fix: MoE Backend defaults (1233)` into `r0.3.0` (#1245)** (6a92418)
- **cp: `fix: fix configs (1209)` into `r0.3.0` (#1246)** (6f3b432)
- **cp: `fix: handle LoRA adapters with fused qkv/mlp (1227)` into `r0.3.0` (#1240)** (441c815)
- **cp: `fix: safetensors for peft (1238)` into `r0.3.0` (#1239)** (2760fa9)
- **cp: `fix: mistral tokenizer (1213)` into `r0.3.0` (#1237)** (c5db441)
- **cp: `fix: Llama3 RoPE fix (1220)` into `r0.3.0` (#1236)** (530588c)
- **cp: `fix: get_dp_group_size for DDP (1211)` into `r0.3.0` (#1234)** (d743260)

### Backend
- **cp: `ci: Update release workflow to include changelog and docs (1320)` into `r0.3.0` (#1366)** (8e81e40)
- **cp: `refactor: extract initialize_model_weights from load_base_model (1356)` into `r0.3.0` (#1360)** (7b8015a)
- **cp: `ci: Update release-docs workflow to use v0.72.0 (1295)` into `r0.3.0` (#1315)** (df3f920)
- **cp: `build: Bump mlflow, urllib3, and fonttools (1334)` into `r0.3.0` (#1344)** (d1de754)
- **cp: `build: Explicitly set torch minimum version as 2.6.0 (1299)` into `r0.3.0` (#1314)** (948cdc7)
- **cp: `resume activation_checkpointing in qwen3_moe_30b_lora (1279) into r0.3.0` (#1310)** (3420aed)
- **cp: chery-pick 1265 in r0.3.0 (#1293)** (d05589a)
- **cp: `feat: passthrough if inputs_embeds passed into nanov3 (1261)` into `r0.3.0` (#1269)** (20a0e9f)
- **cp: `ci: Address cve from rc4 (1253)` into `r0.3.0` (#1257)** (e5121cf)
- **cp: cherry pick #1225 (#1255)** (c9f402b)
- **cp: `ci: Revert L2_Pretrain_and_KD to Azure runners (1230)` into `r0.3.0` (#1231)** (479fc0d)
- **cp: `ci: Update release docs emails-csv to only use env value (1223)` into `r0.3.0` (#1224)** (962d08d)

### Docs
- **docs: Update docs for 0.3.0 to include nightly (#1359)** (fa31028)
- **docs: Update docs for 0.3.0 (#1348)** (38b9f10)

_Recap by [Repo Wrapped](https://repowrapped.com/gh/NVIDIA-NeMo/Automodel?utm_source=github-action)._