mirror of
https://github.com/huggingface/text-generation-inference.git
synced 2025-04-19 22:02:06 +00:00
3c9df21ff8
* Add support for compressed-tensors w8a8 int checkpoints This change adds a loader for w8a8 int checkpoints. One large benefit of int8 support is that the corresponding cutlass matmul kernels also work on compute capability 7.5. Evaluation on neuralmagic/Meta-Llama-3.1-8B-Instruct-quantized.w8a8: | Tasks |Version| Filter |n-shot| Metric | |Value | |Stderr| |---------------|------:|----------------|-----:|-----------------------|---|-----:|---|------| |gsm8k_cot_llama| 3|flexible-extract| 8|exact_match |↑ |0.8431|± |0.0100| | | |strict-match | 8|exact_match |↑ |0.8393|± |0.0101| |ifeval | 4|none | 0|inst_level_loose_acc |↑ |0.8597|± | N/A| | | |none | 0|inst_level_strict_acc |↑ |0.8201|± | N/A| | | |none | 0|prompt_level_loose_acc |↑ |0.7967|± |0.0173| | | |none | 0|prompt_level_strict_acc|↑ |0.7468|± |0.0187| Which is the same ballpark as vLLM. As usual, lots of thanks to Neural Magic/vLLM for the kernels. * Always use dynamic input quantization for w8a8 int It's far less flaky and gives better output. * Use marlin-kernels 0.3.5 * Fix a typo Co-authored-by: drbh <david.richard.holtz@gmail.com> * Small fixes --------- Co-authored-by: drbh <david.richard.holtz@gmail.com>
91 lines
2.4 KiB
Python
91 lines
2.4 KiB
Python
import pytest
|
|
|
|
|
|
@pytest.fixture(scope="module")
|
|
def compressed_tensors_w8a8_int_handle(launcher):
|
|
with launcher(
|
|
"neuralmagic/Llama-3.2-3B-Instruct-quantized.w8a8",
|
|
num_shard=2,
|
|
quantize="compressed-tensors",
|
|
) as handle:
|
|
yield handle
|
|
|
|
|
|
@pytest.fixture(scope="module")
|
|
async def compressed_tensors_w8a8_int(compressed_tensors_w8a8_int_handle):
|
|
await compressed_tensors_w8a8_int_handle.health(300)
|
|
return compressed_tensors_w8a8_int_handle.client
|
|
|
|
|
|
@pytest.mark.release
|
|
@pytest.mark.asyncio
|
|
@pytest.mark.private
|
|
async def test_compressed_tensors_w8a8_int(
|
|
compressed_tensors_w8a8_int, response_snapshot
|
|
):
|
|
response = await compressed_tensors_w8a8_int.generate(
|
|
"What is deep learning?",
|
|
max_new_tokens=10,
|
|
decoder_input_details=True,
|
|
)
|
|
|
|
assert (
|
|
response.generated_text
|
|
== " and how does it differ from traditional machine learning?\n"
|
|
)
|
|
assert response.details.generated_tokens == 10
|
|
assert response == response_snapshot
|
|
|
|
|
|
@pytest.mark.release
|
|
@pytest.mark.asyncio
|
|
@pytest.mark.private
|
|
async def test_compressed_tensors_w8a8_int_all_params(
|
|
compressed_tensors_w8a8_int, response_snapshot
|
|
):
|
|
response = await compressed_tensors_w8a8_int.generate(
|
|
"What is deep learning",
|
|
max_new_tokens=10,
|
|
repetition_penalty=1.2,
|
|
return_full_text=True,
|
|
stop_sequences=["test"],
|
|
temperature=0.5,
|
|
top_p=0.9,
|
|
top_k=10,
|
|
truncate=5,
|
|
typical_p=0.9,
|
|
watermark=True,
|
|
decoder_input_details=True,
|
|
seed=0,
|
|
)
|
|
|
|
assert response.details.generated_tokens == 10
|
|
assert (
|
|
response.generated_text
|
|
== "What is deep learning?\nDeep learning, also known as neural network or"
|
|
)
|
|
assert response == response_snapshot
|
|
|
|
|
|
@pytest.mark.release
|
|
@pytest.mark.asyncio
|
|
@pytest.mark.private
|
|
async def test_compressed_tensors_w8a8_int_load(
|
|
compressed_tensors_w8a8_int, generate_load, response_snapshot
|
|
):
|
|
responses = await generate_load(
|
|
compressed_tensors_w8a8_int,
|
|
"What is deep learning?",
|
|
max_new_tokens=10,
|
|
n=4,
|
|
)
|
|
|
|
assert (
|
|
responses[0].generated_text
|
|
== " and how does it differ from traditional machine learning?\n"
|
|
)
|
|
assert len(responses) == 4
|
|
assert all([r.generated_text == responses[0].generated_text for r in responses])
|
|
|
|
assert responses == response_snapshot
|