mimo-vllm: garble root cause = instanttensor mis-maps MiMo split-GDN weights; b12x loader contract (mm-encoder-tp-mode weights, full Omni, bf16 KV, VLLM_PLUGINS=b12x_loader); dflash re-enabled per Pro launcher; quality bench started

recipes/mimo26-flash-mopd-4x-vllm/recipe.toml

@@ -74,18 +74,26 @@ args = [
     # Mirror of the lab's serve-mimo26-flash.sh: b12x linear + MoE backends.
     "--linear-backend", "b12x",
     "--moe-backend", "b12x",
-    "--load-format", "instanttensor",
+    # 2026-09-28 garble bisect, boot B1: mirror the lab launcher exactly —
+    # the b12x loader (plugin env below), full Omni + --mm-encoder-tp-mode
+    # weights instead of --language-model-only, explicit bf16 KV. instanttensor
+    # was our deviation; MiMo's split-GDN weight mapping is what the b12x
+    # loader path qualifies.
+    "--load-format", "b12x",
+    "--mm-encoder-tp-mode", "weights",
+    "--kv-cache-dtype", "bfloat16",
     "--kernel-config", """
 '{"enable_cutedsl_warmup":false,"enable_flashinfer_autotune":false}'""",
     # DFlash draft from the checkpoint's dflash/ subfolder (7 tokens). The
     # karmic-kraken b12x supports non-causal attention, so the draft runs
     # B12X as well (the TRITON_ATTN workaround was only needed on r39).
+    # DFlash-7 draft, B12X attention, kv_cache_dtype auto — the lab's
+    # Pro-TP8 DFlash launcher contract verbatim. The 2026-09-28 garble was
+    # NOT spec-related: instanttensor mis-mapped the split-GDN weights; the
+    # b12x loader fixed it (see --load-format comment).
     "--speculative-config", """
-'{"method":"dflash","model":"{{dflash}}","num_speculative_tokens":7,"attention_backend":"B12X"}'""",
-    # Nightly resolves the Omni wrapper's eagle3 hook on the inner language
-    # model natively (no image patch needed); --language-model-only keeps
-    # the visual/audio towers out of the memory budget.
-    "--language-model-only",
+'{"method":"dflash","model":"{{dflash}}","num_speculative_tokens":7,"kv_cache_dtype":"auto","attention_backend":"B12X"}'""",
+    # (the --language-model-only line removed for boot B1 — see above)
     "--reasoning-parser", "mimo",
     "--tool-call-parser", "mimo",
     "--enable-auto-tool-choice",
@@ -119,6 +127,7 @@ container = "/root/.cache/vllm"
 
 [env]
 HF_HUB_OFFLINE = "1"
+VLLM_PLUGINS = "b12x_loader"
 VLLM_ENABLE_ROCE_ALLREDUCE = "1"
 MAX_JOBS = "2"
 VLLM_ENGINE_READY_TIMEOUT_S = "3600"

recipes/mimo26-flash-mopd-4x-vllm/reports/2026-09-28_193842_vllm-kk-nightly.json

@@ -0,0 +1,11 @@
+{
+  "schema_version": "1",
+  "tool_eval_bench_version": "2.7.1.dev7+gbd35ba91b",
+  "final_score": null,
+  "rating": null,
+  "safety_warnings": [],
+  "deployability": null,
+  "responsiveness": null,
+  "total_scenarios": null,
+  "error": "[Errno 5] Input/output error"
+}

recipes/mimo26-flash-mopd-4x-vllm/reports/2026-09-28_193842_vllm-kk-nightly.log

@@ -6,3 +6,26 @@
 │ latency=generation                                                           │
 ╰──────────────────────────────────────────────────────────────────────────────╯
 
+  ✓ Complete ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 18/18 0:02:56
+
+  llama-benchy 0.4.0
+  Estimated latency: 106.1 ms
+
+                               llama-benchy Results                             
+┏━━━━━━━━━━━━━━━━━━━━━━┳━━━━┳━━━━━━━━━┳━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━
+┃                      ┃    ┃         ┃         ┃     TTFT ┃    Total ┃         
+┃ Test                 ┃ c  ┃  pp t/s ┃  tg t/s ┃     (ms) ┃     (ms) ┃   Tokens
+┡━━━━━━━━━━━━━━━━━━━━━━╇━━━━╇━━━━━━━━━╇━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━
+│ pp2048 tg128 @ d0    │ c1 │   3,956 │    18.3 │      641 │    7,530 │ 2048+128
+│ pp2048 tg128 @ d0    │ c2 │   3,531 │    28.1 │      871 │    9,403 │ 2048+128
+│ pp2048 tg128 @ d4096 │ c1 │   3,771 │    19.3 │    1,791 │    8,325 │ 2048+128
+│ pp2048 tg128 @ d4096 │ c2 │   3,557 │    23.3 │    2,671 │   12,453 │ 2048+128
+│ pp2048 tg128 @ d8192 │ c1 │   4,379 │    18.9 │    2,450 │    9,134 │ 2048+128
+│ pp2048 tg128 @ d8192 │ c2 │   4,309 │    23.6 │    3,832 │   13,385 │ 2048+128
+└──────────────────────┴────┴─────────┴─────────┴──────────┴──────────┴─────────
+
+  ℹ Metrics sourced from llama-benchy — see https://github.com/eugr/llama-benchy
+for methodology.
+
+Excluding TC-45 from scoring: the endpoint does not enforce tool_choice='required'.
+{"event": "scenario_start", "scenario_id": "TC-01", "title": "Direct Specialist Match", "category": "A", "index": 0, "total": 92}

recipes/mimo26-flash-mopd-4x-vllm/reports/2026-09-28_203145_vllm-kk-nightly-dflash.log

@@ -0,0 +1,8 @@
+^D{"event": "backend_detected", "backend": "vllm"}
+
+╭──────────────────── ⚡ llama-benchy Throughput Benchmark ────────────────────╮
+│ mimo-v2.6-flash                                                              │
+│ pp=[2048]  tg=[128]  depth=[0, 4096, 8192]  concurrency=[1, 2]  runs=3       │
+│ latency=generation                                                           │
+╰──────────────────────────────────────────────────────────────────────────────╯
+