biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Model] New model support for microsoft/Phi-4-mini-flash-reasoning (#20702)

Browse Source 
Signed-off-by: Congcong Chen <congcongchen@microsoft.com>
This commit is contained in:
Congcong Chen
2025-07-12 06:02:10 -07:00
committed by GitHub
parent b639327ad9
commit 2c11a738b3
22 changed files with 1869 additions and 41 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
tests/models/registry.py
View File

@@ -248,6 +248,10 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
"Phi3SmallForCausalLM": _HfExamplesInfo("microsoft/Phi-3-small-8k-instruct",
trust_remote_code=True,
v0_only=True),
"Phi4FlashForCausalLM": _HfExamplesInfo("microsoft/Phi-4-mini-flash-reasoning", # noqa: E501
trust_remote_code=True,
v0_only=True,
max_model_len=10240),
"PhiMoEForCausalLM": _HfExamplesInfo("microsoft/Phi-3.5-MoE-instruct",
trust_remote_code=True),
"Plamo2ForCausalLM": _HfExamplesInfo("pfnet/plamo-2-1b",

3
tests/models/test_initialization.py
View File

@@ -103,6 +103,9 @@ def test_can_initialize(model_arch: str, monkeypatch: pytest.MonkeyPatch):
_initialize_kv_caches_v1), monkeypatch.context() as m):
if model_info.v0_only:
m.setenv("VLLM_USE_V1", "0")
if model_arch == "Phi4FlashForCausalLM":
# Phi4FlashForCausalLM only supports DIFFERENTIAL_FLASH_ATTN backend
m.setenv("VLLM_ATTENTION_BACKEND", "DIFFERENTIAL_FLASH_ATTN")
LLM(
model_info.default,
tokenizer=model_info.tokenizer,

25
tests/test_utils.py
View File

@@ -458,6 +458,31 @@ def test_bind_kv_cache():
assert ctx['layers.2.self_attn'].kv_cache[0] is kv_cache[2]
assert ctx['layers.3.self_attn'].kv_cache[0] is kv_cache[3]
def test_bind_kv_cache_kv_sharing():
from vllm.attention import Attention
ctx = {
'layers.0.self_attn': Attention(32, 128, 0.1),
'layers.1.self_attn': Attention(32, 128, 0.1),
'layers.2.self_attn': Attention(32, 128, 0.1),
'layers.3.self_attn': Attention(32, 128, 0.1),
}
kv_cache = [
torch.zeros((1, )),
torch.zeros((1, )),
torch.zeros((1, )),
torch.zeros((1, )),
]
shared_kv_cache_layers = {
'layers.2.self_attn': 'layers.1.self_attn',
'layers.3.self_attn': 'layers.0.self_attn'
}
bind_kv_cache(ctx, [kv_cache], shared_kv_cache_layers)
assert ctx['layers.0.self_attn'].kv_cache[0] is kv_cache[0]
assert ctx['layers.1.self_attn'].kv_cache[0] is kv_cache[1]
assert ctx['layers.2.self_attn'].kv_cache[0] is kv_cache[1]
assert ctx['layers.3.self_attn'].kv_cache[0] is kv_cache[0]
def test_bind_kv_cache_non_attention():
from vllm.attention import Attention