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

fix(deepseek-v41): coli run, the chat's stderr, and a missing header dependency

JustVugg23f98cd9affe
diff --git a/c/Makefile b/c/Makefile
index c384421d..f21c59ba 100644
--- a/c/Makefile
+++ b/c/Makefile
@@ -1049,7 +1049,7 @@ qwen36$(EXE): qwen36.c serve_poll.h cli_args.h qwen36_tier.h st.h json.h compat.
# stream from the official checkpoint and quant.h's mxfp4 kernel reads them as they
# are, so there is no conversion target here to go with it.
deepseek_v41$(EXE): deepseek_v41.c cli_args.h st.h json.h tok.h tok_unicode.h compat.h quant.h \
- hyper_connections.h serve_codec.h serve_poll.h .build-config
+ sparse_attn.h hyper_connections.h serve_codec.h serve_poll.h .build-config
$(CC) $(CFLAGS) deepseek_v41.c -o deepseek_v41$(EXE) $(LDFLAGS)
qwen38$(EXE): qwen38.c cli_args.h qwen38_core.h qwen38_nfc.h qwen38_nfc_tables.h st.h json.h compat.h quant.h route_trace.h tok.h tok_unicode.h tok_unicode_o200k.h serve_codec.h edge_adapter_internal.h edge_adapters.h edge_runtime.h qwen38_vision.h segment_adapter_internal.h segment_adapters.h segment_runtime.h
diff --git a/c/deepseek_v41.c b/c/deepseek_v41.c
index 21513bca..6f1fcb90 100644
--- a/c/deepseek_v41.c
+++ b/c/deepseek_v41.c
@@ -247,6 +247,15 @@ typedef struct { float *w; int64_t n; } WF; /* f32
#define FP8_TILE 32
+/* Per-turn accounting on stderr. Off by default because `coli chat` shows the
+ * engine's stderr right next to the answer; V41_STATS=1 turns it on, which is
+ * what the benchmarks and the server scripts use. */
+static int v41_stats(void) {
+ static int on = -1;
+ if (on < 0) on = getenv("V41_STATS") ? atoi(getenv("V41_STATS")) : 0;
+ return on;
+}
+
/* V41_TRACE=1 prints a checksum of the tensors the reference prints too, so a
* divergence is located by diffing two columns rather than by reading two forwards. */
static int g_trace = 0;
@@ -2967,10 +2976,11 @@ static int spec_ready(Spec *sp) {
* half of what it needs to be. */
if (sp->window_prop >= 10 &&
sp->window_acc * 100 < sp->window_prop * (uint64_t)sp->min_accept) {
- fprintf(stderr, "[v41] DSpark: %.0f%% of the last %llu drafts accepted (under "
- "%d%%), pausing for 64 tokens\n",
- 100.0 * (double)sp->window_acc / (double)sp->window_prop,
- (unsigned long long)sp->window_prop, sp->min_accept);
+ if (v41_stats())
+ fprintf(stderr, "[v41] DSpark: %.0f%% of the last %llu drafts accepted "
+ "(under %d%%), pausing for 64 tokens\n",
+ 100.0 * (double)sp->window_acc / (double)sp->window_prop,
+ (unsigned long long)sp->window_prop, sp->min_accept);
sp->pause = 64;
sp->window_prop = sp->window_acc = 0;
return 0;
@@ -3228,7 +3238,12 @@ static void serve_loop(Model *m, Tok *tokenizer, const char *snap) {
}
}
free(batch); free(confidence);
- if (m->engram.active) {
+ /* Per-turn accounting goes to stderr, and stderr is what the user reads
+ * their answer next to in `coli chat`: four lines a turn between the
+ * question and the reply is noise, not instrumentation. Off unless asked
+ * for, and the benches ask for it. */
+ int stats = v41_stats();
+ if (stats && m->engram.active) {
/* What the n-gram memory actually cost this turn. Rows are 264
* bytes and the traffic is Zipfian, so the useful question is not
* how many rows were read but how many of them the cache already
@@ -3245,7 +3260,7 @@ static void serve_loop(Model *m, Tok *tokenizer, const char *snap) {
tab->misses * (double)(m->engram.head_dim + m->engram.head_dim / 32) / 1e6);
}
}
- {
+ if (stats) {
/* Expert streaming, per turn: how many bytes the LRU had to fetch and
* what rate the reads actually achieved. The rate is the number that
* matters -- the device is worth several GB/s, so a figure far under
@@ -3259,7 +3274,7 @@ static void serve_loop(Model *m, Tok *tokenizer, const char *snap) {
seconds > 0 ? turn_bytes / 1e9 / seconds : 0.0,
prefill_bytes / 1e6, prefill_disk, prefill_expert, prefill_wall);
}
- if (m->spec.active && m->spec.proposed > proposed0)
+ if (stats && m->spec.active && m->spec.proposed > proposed0)
fprintf(stderr, "[v41] DSpark: %llu of %llu drafts accepted this turn\n",
(unsigned long long)(m->spec.accepted - accepted0),
(unsigned long long)(m->spec.proposed - proposed0));
diff --git a/c/family_registry.py b/c/family_registry.py
index b7ef9ada..e13fb9b1 100644
--- a/c/family_registry.py
+++ b/c/family_registry.py
@@ -1327,7 +1327,13 @@ FAMILIES = (
# at the gateway rather than desync the wire.
capabilities=FamilyCapabilities(True, False, False, True),
has_gateway_adapter=True,
- has_cli_adapter=True,
+ # coli run stays unwired, for the reason qwen36 gives above and one more:
+ # cmd_run dispatches per arch after this gate, and with no deepseek_v41
+ # branch of its own the launcher would fall through to GLM's binary and
+ # GLM's prompt template. The engine speaks the SERVE protocol and nothing
+ # else, so a one-shot has nowhere to go but the gateway -- which is what
+ # coli chat, coli serve and coli web already use.
+ has_cli_adapter=False,
),
)
diff --git a/c/tests/test_dsv41_dspark_serve.py b/c/tests/test_dsv41_dspark_serve.py
index c1430e4f..710e6516 100644
--- a/c/tests/test_dsv41_dspark_serve.py
+++ b/c/tests/test_dsv41_dspark_serve.py
@@ -54,7 +54,10 @@ def has_draft_head():
def completion(drafts):
"""One turn through the gateway, with drafting on or off. Returns (text, log)."""
port = free_port()
- environment = {**os.environ, "V41_DSPARK": "1" if drafts else "0"}
+ # V41_STATS: the per-turn "N of M drafts accepted" line is what this test reads,
+ # and it is off by default so that `coli chat` does not print accounting between
+ # the question and the answer. A test that measures asks to be told.
+ environment = {**os.environ, "V41_DSPARK": "1" if drafts else "0", "V41_STATS": "1"}
process = subprocess.Popen(
[sys.executable, str(HERE / "coli"), "serve", "--model", str(FIXTURE),
"--port", str(port), "--cap", "4"],
diff --git a/docs/deepseek-v41.md b/docs/deepseek-v41.md
index 63b2134e..8b396705 100644
--- a/docs/deepseek-v41.md
+++ b/docs/deepseek-v41.md
@@ -240,6 +240,7 @@ and costs the contention. Anything that tries again has to start from that.
| variable | default | what it does |
|---|---|---|
+| `V41_STATS` | off | per-turn accounting on stderr: the n-gram cache, the expert bytes and their rate, and how many drafts were accepted. Off by default because `coli chat` shows the engine's stderr next to the answer. |
| `V41_READ_DEPTH` | 8 | expert tensors read at once when a step misses the cache (see above). `1` restores the serial read the engine used to do, for an A/B. |
| `V41_ENGRAM_ROWS` | 65536 | rows of engram cache per table. The traffic is Zipfian: common 2-grams repeat constantly, so a small cache absorbs most of it. 65536 rows is 64 MB per table on the released head_dim. |
| `V41_INDEX_OWNER` | unset | each layer scores against its own owner's index keys (see above). Changes the model's behaviour. |