dankclaw: DSpark parallel crash - ROCm illegal memory access under 2x16k prose, pin PARALLEL=1 per user
This commit is contained in:
@@ -11,17 +11,22 @@ PORT=7770
|
||||
# contextWindow=262144 for frank-default-large. Model is 29GB; ROCm has
|
||||
# ~70GB headroom, so 2x262k quantized KV fits. Verify at startup.
|
||||
CTX_SIZE=524288
|
||||
PARALLEL=2
|
||||
# Parallel slots. 2026-08-15: two concurrent 16k prose prompts (high
|
||||
# prompt-cache reuse) crashed DSpark with 'ROCm error: an illegal memory
|
||||
# access was encountered' inside per-sequence block sampling
|
||||
# (ggml_cuda_synchronize in common_speculative_impl_draft_dflash::draft,
|
||||
# status 6/ABRT, systemd restarted). DSpark multi-slot is NOT safe on this
|
||||
# box, same class as the DFlash shared-draft-context bug. Pinned to 1 slot
|
||||
# per user instruction; revisit after an upstream multi-slot DSpark fix.
|
||||
PARALLEL=1
|
||||
# Spec: DSpark sidecar draft (magnitudedev/Qwen3.8-27B-DSpark-GGUF Q8_0,
|
||||
# 1.36B, block_size 7, Markov confidence head). Switched from bundled-MTP
|
||||
# 2026-08-15 at the user's request. DSpark is the semi-autoregressive
|
||||
# extension of DFlash (DeepSeek, arXiv 2607.05147): parallel diffusion
|
||||
# backbone + sequential Markov head + confidence-scheduled verification.
|
||||
# n_max clamps to the trained block_size 7 (server warning if set higher).
|
||||
# DSpark extends DFlash (which had a shared-draft-context premature-EOS bug)
|
||||
# but adds per-sequence Markov/confidence machinery; parallel safety on this
|
||||
# box is still pending the two-16k-prompt test 2026-08-15. Revert to
|
||||
# bundled-MTP or set PARALLEL=1 if the parallel test fails.
|
||||
# Parallel safety FAILED on this box (see the PARALLEL=1 note above);
|
||||
# single-slot only until an upstream multi-slot DSpark fix lands.
|
||||
# DRAFT_FILE names the sidecar GGUF in MODEL_DIR (downloaded 2026-08-15,
|
||||
# sha256 b007a76b2ce57c1a3ecf36046543aebf60a763861624647d653f16d336c781d2).
|
||||
SPEC_TYPE=draft-dspark
|
||||
|
||||
Reference in New Issue
Block a user