[Bug]: MTP accepts no draft tokens with use_kv_cache_manager_v2=false on DSA models (acceptance length 1.0)

System Info

• GPU: 8x NVIDIA H200 NVL (SM90; two 4-GPU NVLink islands, cross-island over PCIe/UPI), driver 595.71.05
• Host: Ubuntu 24.04.4, 2 TiB RAM, Xeon 6747P
• Container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc25 (CUDA 13.2, Python 3.12.3, torch 2.12.0a0+nv26.5, transformers 5.5.4)
• TensorRT-LLM: 1.3.0rc25 (tag v1.3.0rc25, 785c948); the code paths cited below were also checked on main @ a6616d6f (2026-09-03)
• Backend: PyTorch (trtllm-serve --backend pytorch)

Who can help?

@NVShreyas (GLM-5 / DSA MTP, #11990) @yizhang-nv (DSA cache managers)

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

GLM-5.3 W4AFP8 on the DeepSeek-V3.2 path (DSA), pp_size 2 --tp_size 4 --ep_size 4, speculative_config: {decoding_type: MTP, max_draft_len: 1}, kv_cache_config: {use_kv_cache_manager_v2: false, enable_block_reuse: true} (V1 KVCacheManager + DSACacheManager; startup logs [DSACacheManager] Indexer k-cache: 12 of 42 local layers own an indexer k-cache, Using Sampler: SpecSampler, CUDA graphs captured with draft_len=1).

Serve real multi-turn agentic traffic (or any long prompts) and read /metrics (JSON iteration stats, enable_iter_perf_stats: true):

iters=1000 drafts=3500 accepted=12 gen_req_iters=3500 ctx_req_iters=0 iter_lat_ms p50=42.5 max=81.8

i.e. every generating request gets one draft token per iteration and essentially none is ever accepted (specDecodingStats.acceptanceLength = 1.0 on every iteration; trtllm_spec_decode_acceptance_length = 1.0 in Prometheus). The same checkpoint, same parallelism and same MTP config with the V2 manager (default use_kv_cache_manager_v2: auto) reports avg_decoded_tokens_per_iter ~1.8 in chat responses, so the draft layer itself is fine.

Outputs remain correct with V1 (200/200 replayed requests clean), so the drafts are simply always rejected: MTP costs a draft forward per step and gains nothing.

Expected behavior

Comparable acceptance with V1 and V2 (or a startup warning/error that MTP is unsupported with the V1 DSA cache manager).

actual behavior

Silent zero acceptance with V1; no warning.

additional notes

• V1 was the only manager that survived our long-context replay long enough to measure (V2 failed three ways: #18660, #18661), so users falling back to V1 lose speculative decoding on DSA models without noticing.
• Guess at the mechanism: the MTP layer's DSA indexer state (index_share_for_mtp_iteration) or its indexer K-cache is not mirrored/updated for the draft step under the V1 DSACacheManager, so the draft attends over stale keys. Not verified.

Before submitting a new issue...

• Make sure you already searched for relevant issues, and checked the documentation and examples for answers to frequently asked questions.

Start with the PyTorch MTP path and the DSACacheManager/KVCacheManager interaction, focusing on index_share_for_mtp_iteration and the DSA indexer K-cache during the draft step. Reproduce the V1 configuration with enable_iter_perf_stats and compare /metrics acceptanceLength against V2. Done means V1 achieves comparable acceptance or emits a clear unsupported-configuration warning/error, with regression coverage added where the relevant tests are located.