|
|
@@ -62,54 +62,34 @@ args = [
|
|
|
"--served-model-name", "dsv41",
|
|
|
"--trust-remote-code",
|
|
|
"--distributed-executor-backend", "mp",
|
|
|
- # GB10 startup-gate value (see header); KV pool is engine-computed from
|
|
|
- # the remainder (no pin — their profile pins nothing either).
|
|
|
- "--gpu-memory-utilization", "0.88",
|
|
|
- # 512K context (our floor — their qualified 131K is not usable for us;
|
|
|
- # the KV pool is engine-computed from the memory budget, unchanged):
|
|
|
+ # Bisect C (2026-09-14 morning typo hunt): the FULL PROD CONTRACT on the
|
|
|
+ # r37 image - prod's memory/compute flags (GMU 0.85, seqs 8, tokens 8192,
|
|
|
+ # block 128, K5-tuned capture sizes), no B12X attention/linear backends,
|
|
|
+ # no instanttensor, prod's K5/block spec. Bisect log:
|
|
|
+ # A: removed --linear-backend b12x -> typos PERSIST (not the cause)
|
|
|
+ # B: prod's K5 spec instead of K7 greedy/adaptive -> typos PERSIST
|
|
|
+ # C: if clean -> a contract flag is the cause (re-add one by one);
|
|
|
+ # if garbled -> the r37 tree's DeepSeek model code is the cause
|
|
|
+ # (file with the lab; the r37 tree is not qualified for DS4.1).
|
|
|
+ "--gpu-memory-utilization", "0.85",
|
|
|
"--max-model-len", "524288",
|
|
|
- "--decode-context-parallel-size", "1",
|
|
|
- "--max-num-seqs", "4",
|
|
|
- "--max-num-batched-tokens", "4096",
|
|
|
- "--prefill-compute-share", "0.4",
|
|
|
- "--block-size", "256",
|
|
|
- "--enable-chunked-prefill",
|
|
|
- "--enable-prefix-caching",
|
|
|
- "--kv-cache-dtype", "fp8",
|
|
|
- "--attention-backend", "B12X",
|
|
|
- # 2026-09-14 (typo bisect, in progress): removed --linear-backend b12x -
|
|
|
- # the only flag present in this lane but absent from every clean-text
|
|
|
- # lane (prod: b12x-MoE only; GLM r35: B12X attention, default linear).
|
|
|
- # Hypothesis: the B12X dense-GEMM path (auto activation mode) corrupts
|
|
|
- # close-token logits. If the typo test is clean without it, this stays
|
|
|
- # removed and the cause is documented in the history.
|
|
|
- "--moe-backend", "b12x",
|
|
|
+ "--max-num-seqs", "8",
|
|
|
+ "--max-num-batched-tokens", "8192",
|
|
|
+ "--block-size", "128",
|
|
|
"--compilation-config", """
|
|
|
-'{"cudagraph_mode":"FULL_AND_PIECEWISE"}'""",
|
|
|
- "--max_cudagraph_capture_size", "128",
|
|
|
- # Native DSpark K7 — the draft weights live inside the target checkpoint
|
|
|
- # (no separate draft repo). Greedy draft, standard rejection, adaptive
|
|
|
- # verification (the lab's r36 qualified trio vs our prod's K5
|
|
|
- # probabilistic/block).
|
|
|
+'{"cudagraph_mode":"FULL_AND_PIECEWISE","cudagraph_capture_sizes":[5,6,10,12,15,18,20,24,25,30,35,36,40,42,48]}'""",
|
|
|
"--speculative-config", """
|
|
|
-'{"method":"dspark","num_speculative_tokens":7,"draft_tensor_parallel_size":4,"attention_backend":"B12X","draft_sample_method":"greedy","rejection_sample_method":"standard","enable_adaptive_verification":true}'""",
|
|
|
- # Engram: the model's 196B learned ngram lookup storage. disk mode =
|
|
|
- # io_uring reads of the table rows from the checkpoint dir (bounded
|
|
|
- # staging; NOT pinned RAM). cpu_offload stays false (no generic CPU
|
|
|
- # model offload).
|
|
|
- "--engram-config", """
|
|
|
-'{"cpu_offload":false,"table_memory":"disk"}'""",
|
|
|
- "--load-format", "instanttensor",
|
|
|
- "--safetensors-load-strategy", "lazy",
|
|
|
- "--generation-config", "vllm",
|
|
|
- "--override-generation-config", """
|
|
|
-'{"temperature":1.0,"top_p":0.95}'""",
|
|
|
- "--limit-mm-per-prompt", """
|
|
|
-'{"image":2}'""",
|
|
|
- "--reasoning-parser", "deepseek_v41",
|
|
|
+'{"method":"dspark","num_speculative_tokens":5,"draft_sample_method":"probabilistic","rejection_sample_method":"block","enable_adaptive_verification":false}'""",
|
|
|
+ "--moe-backend", "b12x",
|
|
|
+ # Bisect D: engram disabled (no --engram-config). If the boot fails
|
|
|
+ # (model requires engram) the numeric-tree hypothesis stands; if it boots
|
|
|
+ # and the text is clean -> the lab's engram disk-read path is the cause.
|
|
|
+ # (A/B/C above: all garbled - flags are not the cause.)
|
|
|
"--tool-call-parser", "deepseek_v41",
|
|
|
"--enable-auto-tool-choice",
|
|
|
- "--jit-monitor-mode", "error",
|
|
|
+ "--reasoning-parser", "deepseek_v41",
|
|
|
+ "--default-chat-template-kwargs", """
|
|
|
+'{"thinking": true, "reasoning_effort": "max"}'""",
|
|
|
]
|
|
|
|
|
|
[env]
|
