[dependencies] Upgrade transformers to >=5.0.0,<=5.3.0 by erictang000 · Pull Request #1426 · NovaSky-AI/SkyRL

erictang000 · 2026-04-01T23:02:36Z

Upgrade to transformers v5

Summary

Upgrades transformers from >=4.56.1,<5 to >=5.0.0,<=5.3.0 and adapts SkyRL's model initialization, FSDP loading, and test code to accommodate v5 breaking changes.

CI

Round 2 CI: https://github.com/NovaSky-AI/SkyRL/actions/runs/23917102581 -> 10 failing from before
Megatron CI Round 2: https://github.com/NovaSky-AI/SkyRL/actions/runs/23959241150/job/69884903884 -> 1 failing from before

~~Round 1 CI: https://github.com/NovaSky-AI/SkyRL/actions/runs/23876002482 ~~ -> 17 still failing
~~Megatron CI: https://github.com/NovaSky-AI/SkyRL/actions/runs/23920479124~~

Key changes

Meta-device model initialization (`fsdp_utils.py`, `model_wrapper.py`, `fsdp_worker.py`)

v5 disallows from_pretrained() inside accelerate.init_empty_weights() (TypeError: Parameter.__new__() got an unexpected keyword argument '_is_hf_initialized'). Replaced with:

Rank 0: from_pretrained() (loads real weights)
Non-rank-0: from_config() inside torch.device("meta") (empty shell; weights broadcast by FSDP)

rope_scaling, rope_theta, and _attn_implementation are applied to the config before the branch so both paths are consistent.

FSDP2 non-persistent buffer sync (`fsdp_utils.py`)

from_config on meta produces non-persistent buffers (inv_freq in RotaryEmbedding) with no data. These are excluded from state_dict() and never broadcast. Fixes:

_sync_non_persistent_buffers() broadcasts these from rank 0 after state dict loading
offload_fsdp2_model_to_cpu() now materializes only meta buffers instead of calling model.to_empty() (which wiped all loaded parameters → NaN)

CriticModel `post_init()` (`model_wrapper.py`)

v5 added all_tied_weights_keys in PreTrainedModel.post_init(). The dynamic CriticModel class now calls self.post_init(), and the meta-init path wraps construction in no_init_weights().

Strict dataclass configs (`configs.py`)

PretrainedConfig is now a strict dataclass. Made ModelConfig.__init__ args optional with defaults; fixed get_text_config() signature for v5.

VLM `mm_token_type_ids` (`model_wrapper.py`, VLM tests)

v5 requires mm_token_type_ids for M-RoPE in multimodal models. Threaded through HFModelWrapper.forward() and tests.

Megatron `rope_theta` (`megatron_worker.py`)

v5 moved rope_theta into rope_parameters dict. Added workaround to set provider.rotary_base from the new location.

Other fixes

cuda_ipc_strategy.py: .view(-1) → .reshape(-1) for non-contiguous weight tensors
vllm_server.py: guard sock.close() against uvloop TransportSocket AttributeError
test_remote_inference_client_chat_template.py: use render_chat_completion() for prompt token verification

…sformers_v5

pcmoritz · 2026-04-03T21:02:21Z

        mhc_expansion_rate: mHC expansion rate. Connectors are trainable when this is > 1.
    """

-    # Type hints for config attributes


Do we need to remove these? It would be good to keep them for documentation purposes if possible :)

pcmoritz · 2026-04-03T21:08:42Z

For the tx backend, you will also need to adapt for the change that rope_theta moved to rope_arameters, currently it is failing with

AttributeError: 'ModelConfig' object has no attribute 'rope_theta'

pcmoritz · 2026-04-03T21:13:41Z

+    # Broadcast non-persistent buffers (e.g. inv_freq from RotaryEmbedding) that
+    # are excluded from state_dict.  On non-rank-0 meta-init these are still on
+    # meta device with no data; rank 0 has the correctly computed values.
+    _sync_non_persistent_buffers(model, sharded_sd)


I'm curious do you know why upgrading transformers necessitates this change? Seems a little surprising :)

This was the claude writeup of why this was needed:

Thanks for sharing :)

pcmoritz · 2026-04-03T21:17:15Z

        if hasattr(provider, "q_lora_rank") and hasattr(hf_config, "q_lora_rank"):
            provider.q_lora_rank = hf_config.q_lora_rank

+        # Workaround for transformers v5 moving rope_theta into rope_parameters


Curious why this is needed, since megatron-bridge already updated NVIDIA-NeMo/Megatron-Bridge#2068 -- if this is still needed, should we raise an issue against megatron-bridge so we can remove this workaround going forward?

i believe bumping to latest main on megatron-bridge which i'm planning to do here: #1425

should fix this, but i wanted to isolate the transformers v5 update in this PR.

I can look into removing this in #1425

pcmoritz · 2026-04-03T21:29:25Z

    def __init__(
        self,
-        config: PretrainedConfig | dict,
+        config: PretrainedConfig | dict | None = None,


Do you know why this is needed now? Who is calling this without passing in a config? I'm also concerned that the defaults in

max_lora_adapters: int = 0, max_lora_rank: int = 0, shard_attention_heads: bool = True,

could cause trouble and it would be better to not need to have the **kwargs part, since it can mask problems.

seems like it's this PR in transformers 5.4.0: huggingface/transformers#41250

i'm pinning to <= 5.3.0 so it actually isn't an issue right now (but i guess i was testing with 5.4.0 when originally changing this code). I can revert the changes here for now and we can revisit when upgrading to >=5.4.0.

seems like megatron-bridge caps at <=5.3.0 as well and there's some relevant activity on transformers so these changes could be avoided in the changes: huggingface/transformers#45070

…sformers_v5

erictang000 · 2026-04-04T00:35:16Z

    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
    hf_model = AutoModelForCausalLM.from_pretrained(
-        model_name, attn_implementation="eager", use_safetensors=True, trust_remote_code=True
+        model_name, attn_implementation="eager", use_safetensors=True, torch_dtype=torch.float32


this is needed now since behavior in v5 changed from defaulting to float32 to defaulting to the default model dtype.

SumanthRH

Stamp

The split invocation with `--with transformers==5.2.0` was added in NovaSky-AI#1228 when pyproject.toml still pinned transformers <5, to let the new Qwen 3.5 test use v5 while the rest of the suite stayed on v4. The project-wide migration to v5 in NovaSky-AI#1426 left this carve-out and its comment behind, so test_qwen3_5.py has been artificially pinned to 5.2.0 while everything else runs on whatever pyproject.toml resolves (now 5.5.4 on this branch). Collapse back to a single pytest invocation — the exact shape the workflow had before NovaSky-AI#1228 — so all tests run on one transformers version. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

x

5c18a9a

This comment was marked as resolved.

Sign in to view

x

a6e1d12

This comment was marked as resolved.

Sign in to view

erictang000 added 7 commits April 2, 2026 17:52

x

204988c

x

087e1be

x

e7c3008

x

af9b1de

broadcase non-persistent buffers for meta inint

c91c976

fix megatron CI

e9a5317

Merge branch 'main' of https://github.com/erictang000/SkyRL into tran…

1489733

…sformers_v5

erictang000 requested review from SumanthRH and pcmoritz April 3, 2026 20:38

pcmoritz reviewed Apr 3, 2026

View reviewed changes

erictang000 added 2 commits April 3, 2026 22:57

fix jax tests

996ee9b

more v5 fixes

a0f4c99

This comment was marked as resolved.

Sign in to view

erictang000 added 2 commits April 4, 2026 00:20

Merge branch 'main' of https://github.com/erictang000/SkyRL into tran…

06c18f9

…sformers_v5

x

fb6c861

erictang000 commented Apr 4, 2026

View reviewed changes

pcmoritz approved these changes Apr 4, 2026

View reviewed changes

x

6e29207

This comment was marked as resolved.

Sign in to view

SumanthRH approved these changes Apr 4, 2026

View reviewed changes

fix deepseek rope_scaling and qwen3 routing test failure for v5

18e6731

erictang000 merged commit bcf680f into main Apr 4, 2026
6 checks passed

erictang000 deleted the transformers_v5 branch April 4, 2026 19:42

jamesbraza mentioned this pull request Apr 14, 2026

Gemma 4 support via transformers==5.5 #1509

Open

Conversation

erictang000 commented Apr 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Upgrade to transformers v5

Summary

CI

Key changes

Meta-device model initialization (fsdp_utils.py, model_wrapper.py, fsdp_worker.py)

FSDP2 non-persistent buffer sync (fsdp_utils.py)

CriticModel post_init() (model_wrapper.py)

Strict dataclass configs (configs.py)

VLM mm_token_type_ids (model_wrapper.py, VLM tests)

Megatron rope_theta (megatron_worker.py)

Other fixes

Uh oh!

This comment was marked as resolved.

Uh oh!

This comment was marked as resolved.

Uh oh!

This comment was marked as resolved.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

pcmoritz commented Apr 3, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

This comment was marked as resolved.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

This comment was marked as resolved.

Uh oh!

SumanthRH left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

erictang000 commented Apr 1, 2026 •

edited

Loading

Meta-device model initialization (`fsdp_utils.py`, `model_wrapper.py`, `fsdp_worker.py`)

FSDP2 non-persistent buffer sync (`fsdp_utils.py`)

CriticModel `post_init()` (`model_wrapper.py`)

Strict dataclass configs (`configs.py`)

VLM `mm_token_type_ids` (`model_wrapper.py`, VLM tests)

Megatron `rope_theta` (`megatron_worker.py`)