Skip to content

Commit

Permalink
[Model] Adding support for MSFT Phi-3.5-MoE (vllm-project#7729)
Browse files Browse the repository at this point in the history 
Co-authored-by: Your Name <you@example.com>
Co-authored-by: Zeqi Lin <zelin@microsoft.com>
Co-authored-by: Zeqi Lin <Zeqi.Lin@microsoft.com>
  • Loading branch information
@wenxcs @zelinms
@linzeqipku @dsikka
4 people authored and dsikka committed Sep 5, 2024
1 parent 160feda commit 0907355
Show file tree
Hide file tree
Showing 13 changed files with 1,366 additions and 94 deletions.
4 changes: 4 additions & 0 deletions docs/source/models/supported_models.rst
View file
Original file line number Diff line number Diff line change
Expand Up @@ -147,6 +147,10 @@ Decoder-only Language Models
- Phi-3-Small
- :code:`microsoft/Phi-3-small-8k-instruct`, :code:`microsoft/Phi-3-small-128k-instruct`, etc.
-
* - :code:`PhiMoEForCausalLM`
- Phi-3.5-MoE
- :code:`microsoft/Phi-3.5-MoE-instruct`, etc.
-
* - :code:`PersimmonForCausalLM`
- Persimmon
- :code:`adept/persimmon-8b-base`, :code:`adept/persimmon-8b-chat`, etc.
Expand Down
111 changes: 111 additions & 0 deletions tests/models/test_phimoe.py
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,111 @@
"""Compare the outputs of HF and vLLM for moe models using greedy sampling.
Run `pytest tests/models/test_phimoe.py`.
"""
import pytest
import torch

from vllm.utils import is_cpu

from .utils import check_logprobs_close

MODELS = [
"microsoft/Phi-3.5-MoE-instruct",
]


def test_phimoe_routing_function():
from vllm.model_executor.models.phimoe import phimoe_routing_function
test_case = {
0: {
"hidden_states":
torch.tensor([1, 2, 3, 4, 5, 6, 7, 8],
dtype=torch.float32,
requires_grad=False).view(4, 2),
"gating_output":
torch.tensor([0.1, 0.2, 0.3, 0.4],
dtype=torch.float32,
requires_grad=False),
"topk":
2,
"renormalize":
False,
},
1: {
"hidden_states":
torch.tensor([1, 2, 3, 4, 5, 6, 7, 8],
dtype=torch.float32,
requires_grad=False).view(4, 2),
"gating_output":
torch.tensor([0.4, 0.2, 0.3, 0.4],
dtype=torch.float32,
requires_grad=False),
"topk":
2,
"renormalize":
False,
}
}

ground_truth = {
0: {
"topk_weights":
torch.tensor([1., 1.], dtype=torch.float32, requires_grad=False),
"topk_ids":
torch.tensor([3, 2], dtype=torch.long, requires_grad=False),
},
1: {
"topk_weights":
torch.tensor([0.5, 1.], dtype=torch.float32, requires_grad=False),
"topk_ids":
torch.tensor([0, 3], dtype=torch.long, requires_grad=False),
}
}

for test_id in test_case:
topk_weights, topk_ids = phimoe_routing_function(**test_case[test_id])
assert torch.allclose(topk_weights,
ground_truth[test_id]["topk_weights"])
assert torch.equal(topk_ids, ground_truth[test_id]["topk_ids"])


def get_gpu_memory():
try:
props = torch.cuda.get_device_properties(torch.cuda.current_device())
gpu_memory = props.total_memory / (1024**3)
return gpu_memory
except Exception:
return 0


@pytest.mark.skipif(condition=is_cpu(),
reason="This test takes a lot time to run on CPU, "
"and vllm CI's disk space is not enough for this model.")
@pytest.mark.skipif(condition=get_gpu_memory() < 100,
reason="Skip this test if GPU memory is insufficient.")
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["bfloat16"])
@pytest.mark.parametrize("max_tokens", [64])
@pytest.mark.parametrize("num_logprobs", [5])
def test_models(
hf_runner,
vllm_runner,
example_prompts,
model: str,
dtype: str,
max_tokens: int,
num_logprobs: int,
) -> None:
with hf_runner(model, dtype=dtype) as hf_model:
hf_outputs = hf_model.generate_greedy_logprobs_limit(
example_prompts, max_tokens, num_logprobs)

with vllm_runner(model, dtype=dtype) as vllm_model:
vllm_outputs = vllm_model.generate_greedy_logprobs(
example_prompts, max_tokens, num_logprobs)
check_logprobs_close(
outputs_0_lst=hf_outputs,
outputs_1_lst=vllm_outputs,
name_0="hf",
name_1="vllm",
)
130 changes: 130 additions & 0 deletions ...ayers/fused_moe/configs/E=16,N=3200,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,130 @@
{
"3328": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 2
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"3072": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 2
},
"256": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"num_warps": 4,
"num_stages": 4
},
"768": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"num_warps": 4,
"num_stages": 4
},
"1792": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 4
},
"2560": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 2
},
"2816": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 4
},
"3584": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 2
},
"1536": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 2
},
"2048": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 2
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"num_warps": 4,
"num_stages": 4
},
"3840": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 4
},
"1280": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 2
},
"2304": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 2
},
"4096": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 2
}
}
130 changes: 130 additions & 0 deletions ...ayers/fused_moe/configs/E=16,N=6400,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,130 @@
{
"3840": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"num_warps": 4,
"num_stages": 4
},
"1792": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"num_warps": 4,
"num_stages": 4
},
"3584": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 2
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 2
},
"3072": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 2
},
"2048": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 2
},
"2816": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 4
},
"1280": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 2
},
"768": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"num_warps": 4,
"num_stages": 4
},
"3328": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 2
},
"2560": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"num_warps": 4,
"num_stages": 4
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 8,
"num_warps": 4,
"num_stages": 4
},
"2304": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 2
},
"1536": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 2
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
}
}
Loading

0 comments on commit 0907355

Please sign in to comment.