Skip to content

Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦

C

qwen38 MTP: gated in CI; the harness leaves the test_*.py glob

JustVugg4ac9ed70fa30
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 6dad14ae..5546e466 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -1035,6 +1035,8 @@ jobs:
run: make -C c qwen38-tiny-fp8-check qwen38-tier-engine-check
- name: Routed experts as int4-g64 from the converter's sidecar
run: make -C c qwen38-tiny-int4-check
+ - name: MTP speculative decoding against plain decoding (lossless gate)
+ run: make -C c qwen38-tiny-mtp-check
- name: Repeat the eviction path under ASan and UBSan
run: |
cd c
@@ -1052,6 +1054,10 @@ jobs:
fi
grep -q "Matching tokens: 8/8" qwen38-san.log
grep -q "Matching tokens: 8/8" qwen38-san-int4.log
+ # the MTP head, the S=2 verify and the rollback on every fixture kind
+ ASAN_OPTIONS=detect_leaks=0 UBSAN_OPTIONS=print_stacktrace=1:halt_on_error=1 \
+ python3 tests/qwen38_mtp_harness.py --engine ./qwen38 --bf16 ./qwen38_tiny_mtp \
+ --fp8 ./qwen38_tiny_fp8_mtp --int4 ./qwen38_tiny_int4_mtp
- name: Refuse config and tensor-shape disagreement
run: |
cd c
diff --git a/c/Makefile b/c/Makefile
index dd2957c8..adf7ac3d 100644
--- a/c/Makefile
+++ b/c/Makefile
@@ -1394,7 +1394,7 @@ qwen38-tiny-int4-check: qwen38-tiny-int4-generate qwen38-tiny-fp8-generate qwen3
# "the MTP head"). The fixtures are the three above plus an MTP head (--mtp:
# random head weights from their own seed under the release's mtp.* names; the
# model and its references do not change) and a byte tokenizer for the text
-# and serve gates. tests/test_qwen38_mtp.py checks:
+# and serve gates. tests/qwen38_mtp_harness.py checks:
# (a) the head's draft logits equal tools/qwen38_mtp_ref.py's float32
# reference for the default wiring b and the diagnostic a, through the
# one-row and the two-row (after an accepted draft) paths, on BF16 and
@@ -1416,7 +1416,7 @@ qwen38-tiny-mtp-generate:
$(PYTHON) tools/make_edge_tiny_tokenizer.py --vocab-size 64 ./$$d || exit $$?; done
qwen38-tiny-mtp-check: qwen38-tiny-mtp-generate qwen38$(EXE)
- $(PYTHON) tests/test_qwen38_mtp.py --engine ./qwen38$(EXE) --bf16 ./qwen38_tiny_mtp \
+ $(PYTHON) tests/qwen38_mtp_harness.py --engine ./qwen38$(EXE) --bf16 ./qwen38_tiny_mtp \
--fp8 ./qwen38_tiny_fp8_mtp --int4 ./qwen38_tiny_int4_mtp
# Context-size gates: KV layout, growth across requests, and the attention
diff --git a/c/tests/test_qwen38_mtp.py b/c/tests/qwen38_mtp_harness.py
similarity index 99%
rename from c/tests/test_qwen38_mtp.py
rename to c/tests/qwen38_mtp_harness.py
index 2e8f748e..6e338940 100644
--- a/c/tests/test_qwen38_mtp.py
+++ b/c/tests/qwen38_mtp_harness.py
@@ -21,7 +21,7 @@ serve). Gates:
serve a multi-turn serve session (prefix reuse, a pinned prefix,
logprobs, a sampled turn) gives the same frames MTP on and off.
- python tests/test_qwen38_mtp.py --engine ./qwen38 --bf16 ./qwen38_tiny_mtp \\
+ python tests/qwen38_mtp_harness.py --engine ./qwen38 --bf16 ./qwen38_tiny_mtp \\
--fp8 ./qwen38_tiny_fp8_mtp --int4 ./qwen38_tiny_int4_mtp
"""
import argparse