Skip to content

README.md

Latest commit

 

History

History
270 lines (193 loc) · 10.1 KB

README.md

File metadata and controls

270 lines (193 loc) · 10.1 KB

Qwen2.5 / Qwen3 Optimized with NVIDIA TransformerEngine

This folder contains source code and tests for Qwen2.5 and Qwen3 style models that inherit from the transformers PreTrainedModel class and use TransformerEngine layers. Users can convert existing Qwen checkpoints from HuggingFace using the provided conversion utilities.

Feature support

The Qwen implementations natively support the following TransformerEngine-provided optimizations:

Feature Support
FP8 Supported on compute capacity 9.0 and above (Hopper+)
MXFP8 Supported on compute capacity 10.0 and 10.3 (Blackwell), 12.0 support pending
Sequence Packing / THD input format Supported
FP8 with THD input format Supported where FP8 is supported
Import from HuggingFace checkpoints Supported
Export to HuggingFace checkpoints Supported
KV-cache inference Supported (including beam search)

Inference Examples

Quick start: convert and run (Qwen3)

Note: The snippets below use bare imports (e.g., from convert_qwen3 import ...). Run them from the bionemo-recipes/models/qwen directory, or install dependencies first with pip install -r requirements.txt.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

from convert_qwen3 import convert_qwen3_hf_to_te

# Load the original HuggingFace Qwen3 model
model_hf = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3-0.6B", torch_dtype=torch.bfloat16
)

# Convert to TransformerEngine
model_te = convert_qwen3_hf_to_te(model_hf)
model_te.to("cuda")

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-0.6B")
tokenizer.pad_token = tokenizer.eos_token

inputs = tokenizer("The quick brown fox", return_tensors="pt")
inputs = {k: v.to("cuda") for k, v in inputs.items()}

with torch.no_grad():
    output_ids = model_te.generate(**inputs, max_new_tokens=16, use_cache=False)

print(tokenizer.decode(output_ids[0], skip_special_tokens=True))

Quick start: convert and run (Qwen2.5)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

from convert_qwen2 import convert_qwen2_hf_to_te

# Load the original HuggingFace Qwen2.5 model
model_hf = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-0.5B-Instruct", torch_dtype=torch.bfloat16
)

# Convert to TransformerEngine
model_te = convert_qwen2_hf_to_te(model_hf)
model_te.to("cuda")

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-0.5B-Instruct")
tokenizer.pad_token = tokenizer.eos_token

inputs = tokenizer("The quick brown fox", return_tensors="pt")
inputs = {k: v.to("cuda") for k, v in inputs.items()}

with torch.no_grad():
    output_ids = model_te.generate(**inputs, max_new_tokens=16, use_cache=False)

print(tokenizer.decode(output_ids[0], skip_special_tokens=True))

Running with Low Precision (FP8/FP4)

The TE-optimized Qwen models support per-layer quantization via two mechanisms: a config-level layer_precision list that declares which layers use which precision, and constructor-level recipe objects (fp8_recipe, fp4_recipe) that control the quantization behaviour.

Configuration: layer_precision

NVQwen2Config.layer_precision (and NVQwen3Config.layer_precision) is a list of length num_hidden_layers where each element is "fp8", "fp4", or None (BF16 fallback). When set, it controls the te.autocast context used for each transformer layer during both initialization and forward pass.

from modeling_qwen3_te import NVQwen3Config, NVQwen3ForCausalLM

# All layers in FP8
config = NVQwen3Config.from_pretrained(
    "Qwen/Qwen3-0.6B",
    layer_precision=["fp8"] * 28,
)

If you pass an fp8_recipe to the model constructor without setting layer_precision, it defaults to ["fp8"] * num_hidden_layers (all layers FP8). You can also mix precisions, for example running most layers in FP8 but keeping the first and last layers in BF16:

layer_precision = [None] + ["fp8"] * 26 + [None]
config = NVQwen3Config.from_pretrained(
    "Qwen/Qwen3-0.6B",
    layer_precision=layer_precision,
)

Constructor arguments: fp8_recipe and fp4_recipe

The model classes (NVQwen2Model, NVQwen2ForCausalLM, NVQwen3Model, NVQwen3ForCausalLM) accept fp8_recipe and fp4_recipe keyword arguments. These are transformer_engine.common.recipe.Recipe objects that configure the quantization algorithm (e.g., delayed scaling, block scaling, MXFP8).

import transformer_engine.common.recipe as te_recipe

from modeling_qwen3_te import NVQwen3Config, NVQwen3ForCausalLM

fp8_recipe = te_recipe.DelayedScaling()

config = NVQwen3Config.from_pretrained(
    "Qwen/Qwen3-0.6B",
    layer_precision=["fp8"] * 28,
)
model = NVQwen3ForCausalLM(config, fp8_recipe=fp8_recipe)

For FP4 (NVFP4) quantization, pass an fp4_recipe instead and set the corresponding layers to "fp4" in layer_precision:

fp4_recipe = te_recipe.NVFP4BlockScaling()

config = NVQwen3Config.from_pretrained(
    "Qwen/Qwen3-0.6B",
    layer_precision=["fp4"] * 28,
)
model = NVQwen3ForCausalLM(config, fp4_recipe=fp4_recipe)

You can also mix FP8 and FP4 layers by providing both recipes and a mixed layer_precision list.

The same pattern applies to Qwen2.5 models using NVQwen2Config and NVQwen2ForCausalLM.

Quantized model initialization: use_quantized_model_init

When use_quantized_model_init=True is set in the config, layers are created inside a te.quantized_model_init context. This tells TransformerEngine to initialize weights directly in the target quantized format, avoiding a separate quantization step after initialization.

config = NVQwen3Config.from_pretrained(
    "Qwen/Qwen3-0.6B",
    layer_precision=["fp4"] * 28,
    use_quantized_model_init=True,
)
model = NVQwen3ForCausalLM(config, fp4_recipe=te_recipe.NVFP4BlockScaling())

Notes

  • The lm_head always runs in higher precision (te.autocast(enabled=False)) regardless of layer_precision, to avoid numerical instability in the output logits.
  • FP8 requires compute capability 9.0+ (Hopper). MXFP8 requires compute capability 10.0+ (Blackwell).
  • If an fp8_recipe is provided without layer_precision, all layers default to FP8. Providing both fp8_recipe and fp4_recipe without layer_precision raises a RuntimeError.
  • An FP4 layer requires an fp4_recipe; omitting it raises a RuntimeError.

Converting Between Model Formats

This section explains how to convert between Hugging Face Transformers and Transformer Engine (TE) Qwen model formats. The process demonstrates bidirectional conversion: from Transformers to TE format for optimized training and inference, and back to Hugging Face Transformers format for sharing and deployment.

Converting from HF Transformers to TE (Qwen3)

from transformers import AutoModelForCausalLM

from convert_qwen3 import convert_qwen3_hf_to_te

model_hf = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-0.6B")
model_te = convert_qwen3_hf_to_te(model_hf)
model_te.save_pretrained("/path/to/te_checkpoint")

Converting from TE back to HF Transformers (Qwen3)

from convert_qwen3 import convert_qwen3_te_to_hf
from modeling_qwen3_te import NVQwen3ForCausalLM

model_te = NVQwen3ForCausalLM.from_pretrained("/path/to/te_checkpoint")
model_hf = convert_qwen3_te_to_hf(model_te)
model_hf.save_pretrained("/path/to/hf_checkpoint")

Converting from HF Transformers to TE (Qwen2.5)

from transformers import AutoModelForCausalLM

from convert_qwen2 import convert_qwen2_hf_to_te

model_hf = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-0.5B-Instruct")
model_te = convert_qwen2_hf_to_te(model_hf)
model_te.save_pretrained("/path/to/te_checkpoint")

Converting from TE back to HF Transformers (Qwen2.5)

from convert_qwen2 import convert_qwen2_te_to_hf
from modeling_qwen2_te import NVQwen2ForCausalLM

model_te = NVQwen2ForCausalLM.from_pretrained("/path/to/te_checkpoint")
model_hf = convert_qwen2_te_to_hf(model_te)
model_hf.save_pretrained("/path/to/hf_checkpoint")

Once converted back to HF format, the model can be loaded by any library that supports Qwen, such as vLLM or SGLang.

Validating Converted Models

To validate the converted models, refer to the commands in Inference Examples above to load and test both the original and converted models to ensure loss and logit values are similar. Additionally, refer to the golden value tests in test_modeling_qwen2_te.py and test_modeling_qwen3_te.py.

Developer Guide

Running tests

To run tests locally, run recipes_local_test.py from the repository root with the model directory as an argument.

./ci/scripts/recipes_local_test.py bionemo-recipes/models/qwen/

Exporting to Hugging Face Hub

The model directory includes an export.py script that bundles all files needed for Hugging Face Hub distribution. To create the export bundle, run from the model directory:

python export.py

Before publishing, validate the export by running the local test suite via recipes_local_test.py.

Development container

To use the provided devcontainer, use "Dev Containers: Reopen in Container" from the VSCode menu, and choose the "BioNeMo Recipes Dev Container" option. To run the tests inside the container, first install the dependencies with pip install -r requirements.txt, then run pytest -v . in the model directory.