Skip to content

Fix KDA equivalence tests and add accelerate dependency #488

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

Merged
jlamypoirier merged 1 commit into from
Apr 23, 2026
+5 −8
Merged

Fix KDA equivalence tests and add accelerate dependency #488

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
12 changes: 4 additions & 8 deletions fast_llm_external_models/tests/test_apriel2/test_mixer_equivalence.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -144,7 +144,7 @@ def kda_mixer_config(kda_config):
"heads": num_heads,
"head_dim": head_dim,
"convolution_layer": {"kernel_size": 4},
"normalization": {"epsilon": 1e-5},
"normalization": {"epsilon": 1e-5, "activation": "sigmoid"},
}


Expand Down Expand Up @@ -1088,9 +1088,8 @@ def test_vs_fla(
fla_cache = FLACache()
apriel_cache = Apriel2Cache(make_apriel2_config(kda_hidden_size, kda_mixer_config))

# Force chunk mode for prefill
fla_kda.mode = "chunk"
apriel_kda.mode = "chunk"
# Match Apriel2's mode selection: fused_recurrent for seq_len<=64 in eval
fla_kda.mode = "fused_recurrent"

# ========== PHASE 1: Initial Prefill ==========
prefill_input = hidden_states[:, :prefill_len, :]
Expand Down Expand Up @@ -1125,7 +1124,6 @@ def test_vs_fla(

# ========== PHASE 2: Decode (single tokens) ==========
fla_kda.mode = "fused_recurrent"
apriel_kda.mode = "fused_recurrent"

for i in range(decode_steps):
pos = prefill_len + i
Expand Down Expand Up @@ -1160,9 +1158,7 @@ def test_vs_fla(
)

# ========== PHASE 3: Prefill again (decode→prefill transition) ==========
# FLA KDA correctly uses initial_state in chunk mode, so this should match
fla_kda.mode = "chunk"
apriel_kda.mode = "chunk"
fla_kda.mode = "fused_recurrent"

prefill2_start = prefill_len + decode_steps
prefill2_input = hidden_states[:, prefill2_start : prefill2_start + prefill2_len, :]
Expand Down
1 change: 1 addition & 0 deletions setup.cfg
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -44,6 +44,7 @@ OPTIONAL =
# Huggingface tools
HUGGINGFACE =
transformers>=4.57.3,<5.0.0
accelerate>=1.4.0
hf-transfer>=0.1.9
datasets>=4.4.1
huggingface-hub>=0.36.0
Expand Down
Loading