Remove unnecessary `load_weights` methods - #44589

Merged

vllm-bot merged 57 commits into

vllm-project:mainfrom

hmellor:remove-simple-load-weights

Jun 29, 2026

Commits on Jun 5, 2026

Enable fused linear layers to load themselves
hmellor
committed
88beb3b
Enable GPTQ extra bias skipping in AutoWeightsLoader
hmellor
committed
41f4584
Try it on a couple of simple weight loaders
hmellor
committed
fcd151b
Fix LoRA loading for these two models
hmellor
committed
a9788ab
Delete some more load_weights methods
hmellor
committed
cca665e
Add patterns from `maybe_remap_kv_scale_name` to `QuantizationConfig.get_cache_scale_mapper`
hmellor
committed
e316238
Use new mappings in `AutoWeightsLoader`
hmellor
committed
665ca0c
Remove some more load_weights methods
hmellor
committed
edf67e6

Commits on Jun 11, 2026

Merge remote-tracking branch 'upstream/main' into remove-simple-load-weights
hmellor
committed
698d2a6
Fix `load_weights` methods for fused case
hmellor
committed
b1f1c9d
Fix BaiChuan tests that depend on old behaviour
hmellor
committed
b5a7191
Handle MergedColumnParallelLinear for LoRA too
hmellor
committed
f5383aa
Delete some more load_weights methods
hmellor
committed
82a7a64
Add debug logs while loading
hmellor
committed
c3a316a
Fix late initialised biases
hmellor
committed
c802faa
Fix GPTQ tests
hmellor
committed
88de67d
fix bnb
hmellor
committed
4069aae

Commits on Jun 13, 2026

Merge branch 'main' into remove-simple-load-weights
hmellor
committed
9d69ba6
Make `vllm.model_executor.utils.get_packed_modules_mapping` check `hf_to_vllm_mapper` too
hmellor
committed
872ff37
Fix `WeightsMapper.get_packed_modules_mapping`
hmellor
committed
3b73687
Better `SupportsQuant._maybe_apply_model_mapping`
hmellor
committed
c657d7d
`BitsAndBytesModelLoader` can be simpler now
hmellor
committed
f2d548b
Use `get_packed_modules_mapping` for `get_supported_lora_modules`
hmellor
committed
41e3a9e
Fix test
hmellor
committed
68085a0
tweaks
hmellor
committed
2f56a42
typo
hmellor
committed
d051813
Mapper must present both shard id and weight name as supported packings
hmellor
committed
5c2a354
Fix test
hmellor
committed
b5bdb58
`AutoWeightsLoader` inject packed mappings from mapper at load time
hmellor
committed
dbf02b1
revert lora test changes
hmellor
committed
a97a1a1
Revert quant/lora hacks; `get_packed_modules_mapping` -> `get_unfused_mapper`
hmellor
committed
a33e90e
Add `packed_modules_mapping` attrs back to models
hmellor
committed
4e5c1e3
Revert now unused LoRA things
hmellor
committed
735d725
Revert lora test
hmellor
committed
a7a65a9
tweak diff
hmellor
committed
2aadace
More accurate comment
hmellor
committed
ed3b43c

Commits on Jun 19, 2026

Commits on Jun 26, 2026

Commits on Jun 27, 2026

Make `validate_shard_id` narrow type properly
hmellor
committed
d7619d2
Pass shard_id as a param of the weight instead of str hacks
hmellor
committed
1f62b5b
Apply to models
hmellor
committed
4ab92e2
revert whisper formatting change
hmellor
committed
5fe0c6e
Remove child scan in bnb loading
hmellor
committed
0e78b7c
Bring GLMV4 mapper to the top class
hmellor
committed
48fae6b
propagate mapper from Qwen2Model
hmellor
committed
2426f76
Add Gemma3
hmellor
committed
d36e06e
propagate up for internlm2 too
hmellor
committed
69f7415
supports quant and propagation for other models too
hmellor
committed
29c0149
Merge branch 'main' into remove-simple-load-weights
hmellor
authored
c3491ea

Commits on Jun 29, 2026

Merge branch 'main' into remove-simple-load-weights
hmellor
authored
0e292fe