Skip to content

feat: add Qwen3.8-27B (dense) support #3319

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
samsja wants to merge 1 commit into
base: main
Choose a base branch
from
Open

feat: add Qwen3.8-27B (dense) support #3319

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -53,7 +53,7 @@ With `[model] impl = "auto"` (the default), the trainer selects that custom stac
| GLM-5 (`glm_moe_dsa`) | `zai-org/GLM-5`, `zai-org/GLM-5-FP8` | yes | ✅ | ✅ |
| Qwen3 MoE (`qwen3_moe`) | `Qwen/Qwen3-30B-A3B`, … | yes | ✅ | ✅ |
| Qwen3.5 MoE (`qwen3_5_moe`) | `Qwen/Qwen3.5-35B-A3B`, … | yes | ✅ | ✅ |
| Qwen3 / Qwen3.5 VLMs | see [advanced.md](docs/advanced.md#multimodal-training) (`qwen3_vl`, `qwen3_5`, `qwen3_5_moe`) | MoE only on MoE VLMs | MoE only | ❌ |
| Qwen3 / Qwen3.5 / Qwen3.8 VLMs | see [advanced.md](docs/advanced.md#multimodal-training) (`qwen3_vl`, `qwen3_5`, `qwen3_5_moe`) — Qwen3.8-27B reuses `qwen3_5` | MoE only on MoE VLMs | MoE only | ❌ |
| Poolside Laguna (`laguna`) | `poolside/Laguna-XS.2` | yes | ✅ | ✅ |
| MiniMax M2 (`minimax_m2`) | `MiniMax/MiniMax-M2` | yes | ✅ | ✅ |
| Nemotron H (`nemotron_h`) | `nvidia/Nemotron-3-Nano-30B-A3B`, `nvidia/Nemotron-3-Super-120B-A12B`, … | yes | ✅ | ✅ |
Expand Down
2 changes: 1 addition & 1 deletion deps/renderers
Open in desktop
Submodule renderers updated 37 files
+8 −0 .github/workflows/style.yml
+3 −3 README.md
+20 −7 docs/renderer-config.md
+4 −0 pyproject.toml
+26 −1 renderers/__init__.py
+85 −8 renderers/base.py
+137 −11 renderers/client.py
+320 −24 renderers/configs.py
+1,369 −0 renderers/gemma4.py
+997 −0 renderers/inkling.py
+33 −0 renderers/laguna_s21.py
+88 −21 renderers/laguna_xs2.py
+22 −3 renderers/nemotron3.py
+195 −7 renderers/parsing.py
+20 −1 renderers/qwen3.py
+112 −55 renderers/qwen35.py
+153 −0 renderers/qwen38.py
+12 −0 tests/conftest.py
+4 −0 tests/test_bridge.py
+14 −0 tests/test_build_helpers.py
+178 −20 tests/test_client.py
+234 −0 tests/test_disabled_thinking_stability.py
+473 −0 tests/test_gemma4.py
+183 −0 tests/test_glm_tool_name_validation.py
+523 −0 tests/test_inkling.py
+232 −0 tests/test_laguna_m1.py
+128 −0 tests/test_laguna_s21.py
+221 −11 tests/test_multimodal.py
+16 −9 tests/test_nemotron3_parity.py
+32 −7 tests/test_nemotron3_ultra.py
+58 −0 tests/test_prime_qwen3_parity.py
+144 −0 tests/test_qwen38.py
+20 −0 tests/test_render_ids.py
+49 −2 tests/test_renderer_config.py
+43 −2 tests/test_renderer_config_parity.py
+6 −0 tests/test_roundtrip.py
+161 −185 uv.lock
2 changes: 1 addition & 1 deletion deps/verifiers
Open in desktop
Submodule verifiers updated 414 files
2 changes: 1 addition & 1 deletion docs/advanced.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -72,7 +72,7 @@ The built-in VLM registry covers:

| Family | `model_type` | Vision attr | LM attr |
|---|---|---|---|
| Qwen3.5 | `qwen3_5` | `model.visual` | `model.language_model` |
| Qwen3.5 / Qwen3.8 (dense) | `qwen3_5` | `model.visual` | `model.language_model` |
| Qwen3.5-MoE | `qwen3_5_moe` | `model.visual` | `model.language_model` |

### Enabling VLM Mode
Expand Down
12 changes: 9 additions & 3 deletions src/prime_rl/trainer/model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -488,7 +488,11 @@ def get_model(

is_vlm_training = config.vlm is not None

if "Qwen3.5" in config.name or "qwen3_5" in config.name.lower():
if (
"Qwen3.5" in config.name
or "Qwen3.8" in config.name
or "qwen3_5" in config.name.lower()
):

@cursor cursor Bot Aug 20, 2026

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Missing KL mismatch table

Medium Severity

This PR adds custom-trainer support for Qwen3.8-27B, but it does not include the required mean KL mismatch table across 20 steps on the math environment with batch_size=64 (all entries < 0.015). That validation is required when introducing a new custom model so trainer and inference stay aligned.

Additional Locations (1)
Fix in Cursor Fix in Web

Triggered by project rule: BugBot Instructions

Reviewed by Cursor Bugbot for commit 6e8a5af. Configure here.

_patch_qwen3_5_text_position_ids()
_patch_qwen3_5_moe_conversion_mapping()
_patch_qwen3_5_linear_attn_varlen()
Expand Down Expand Up @@ -523,8 +527,10 @@ def get_model(

_hub_kernels._kernels_enabled = True

# Fallback Qwen3.5 patch detection from loaded config model_type
if getattr(model_config, "model_type", "").startswith("qwen3_5_moe"):
# Fallback Qwen3.5 patch detection from loaded config model_type. Covers
# renamed Qwen3.5 / Qwen3.8 checkpoints (``qwen3_5``, ``qwen3_5_text``,
# ``qwen3_5_moe``) whose name_or_path misses the name-based gate above.
if getattr(model_config, "model_type", "").startswith("qwen3_5"):
_patch_qwen3_5_text_position_ids()
_patch_qwen3_5_moe_conversion_mapping()
_patch_qwen3_5_linear_attn_varlen()
Expand Down