Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 5 additions & 1 deletion docs/hw-gate.md
Original file line number Diff line number Diff line change
Expand Up @@ -144,11 +144,15 @@ The models (`MODEL_TABLE` in the script, files under `~/models`):
| qwen3-0.6b | Qwen3-0.6B-Q4_K_M.gguf | small dense K-quant | 3 | yes |
| zaya1-8b | zaya1-8b-Q4_K_M.gguf | ZAYA1 (16 experts, 1 active per token), K-quant | 2 | yes |
| qwen3.8-27b | qwen38/UD-Q4_K_XL/Qwen3.8-27B-UD-Q4_K_XL.gguf | gated delta net + attention hybrid, Unsloth dynamic mix | 2 | no |
| qwen3-coder-30b-a3b | Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf | qwen3moe: 128 experts, 8 active per token (MUL_MAT_ID with top-8 routing) | 2 | no |
| qwen3.6-35b-a3b | qwen36/Qwen3.6-35B-A3B-UD-IQ4_XS.gguf | qwen35moe: 256 experts, 8 active per token (MUL_MAT_ID with top-8 routing; IQ4_XS / IQ3_S experts) | 2 | no |
| qwen3-4b-iq4xs | Qwen3-4B-IQ4_XS.gguf | an IQ quant (IQ4_XS matmul and GET_ROWS) | 2 | no |
| qwen3-0.6b-kq8 | Qwen3-0.6B-Q4_K_M.gguf, served with `-fa on -ctk q8_0` | a q8_0 K cache with an f16 V cache | 2 | yes |
| qwen3-0.6b-kvq8 | Qwen3-0.6B-Q4_K_M.gguf, served with `-fa on -ctk q8_0 -ctv q8_0` | q8_0 K and V caches | 2 | yes |

The qwen3-coder-30b-a3b row (Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf, 128 experts) is out of the table until
#2494 closes (engine#2501): its multi-token prefill faults the GPU, so every gate run would reproduce that crash
trigger. The row stays in `MODEL_TABLE` as a comment; re-enable it when #2494 closes.

The last two rows give `llama-server` extra arguments (the fifth column of `MODEL_TABLE`). Their kat
needle (about 4.7K tokens) and the 7.1K-token needle decode past 1,024 KV cells, where the add-on's
split decode kernel (`fa_decode_q8_d128`) takes over from its prompt kernel; the short kat cases
Expand Down
9 changes: 6 additions & 3 deletions scripts/hw-gate.sh
Original file line number Diff line number Diff line change
Expand Up @@ -195,8 +195,8 @@ FLAKY_CASES=${HW_GATE_FLAKY:-"MUL_MAT(type_a=f32,type_b=f32,m=1,n=1,k=2048"}

# name | file under $MODELS | kat --repeat | needle (1 = run tests/hrx_long_prompt.py on it)
# | extra llama-server arguments
# The first three are K-quants (ZAYA1-8B routes 1 of 16 experts). The Qwen3-Coder row covers
# MUL_MAT_ID with top-8 of 128 experts, the IQ row the IQ4_XS matmul and GET_ROWS paths. The
# The first three are K-quants (ZAYA1-8B routes 1 of 16 experts). The Qwen3.6-35B-A3B row covers
# MUL_MAT_ID with top-8 of 256 experts (IQ4_XS / IQ3_S experts, qwen35moe), the IQ row the IQ4_XS matmul and GET_ROWS paths. The
# needle runs on a small and a mid-size model.
# The two -kq8/-kvq8 rows serve Qwen3-0.6B with a q8_0 K cache (V f16, `-ctk q8_0` alone) and with
# q8_0 K and V. Their kat needle (~4.7K tokens) and the 7.1K-token needle decode past 1024 KV
Expand All @@ -207,10 +207,13 @@ FLAKY_CASES=${HW_GATE_FLAKY:-"MUL_MAT(type_a=f32,type_b=f32,m=1,n=1,k=2048"}
# 0..31 only (test-backend-ops has no FLASH_ATTN_EXT case with K and V of different types).
# Rows with server arguments run kat, needle and placement, not kld: the KLD bases are stored per
# model file and recorded with the default f16 cache.
# engine#2501: qwen3-coder-30b-a3b|Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf|2|0 -- re-enable when #2494 closes. Its
# multi-token prefill faults the GPU (MES REMOVE_QUEUE -> MODE2 reset), so every gate run would reproduce the crash
# trigger, and its GGUF was removed from ~/models. Its MoE coverage moved to the qwen3.6-35b-a3b row.
MODEL_TABLE="qwen3-0.6b|Qwen3-0.6B-Q4_K_M.gguf|3|1
zaya1-8b|zaya1-8b-Q4_K_M.gguf|2|1
qwen3.8-27b|qwen38/UD-Q4_K_XL/Qwen3.8-27B-UD-Q4_K_XL.gguf|2|0
qwen3-coder-30b-a3b|Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf|2|0
qwen3.6-35b-a3b|qwen36/Qwen3.6-35B-A3B-UD-IQ4_XS.gguf|2|0
qwen3-4b-iq4xs|Qwen3-4B-IQ4_XS.gguf|2|0
qwen3-0.6b-kq8|Qwen3-0.6B-Q4_K_M.gguf|2|1|-fa on -ctk q8_0
qwen3-0.6b-kvq8|Qwen3-0.6B-Q4_K_M.gguf|2|1|-fa on -ctk q8_0 -ctv q8_0"
Expand Down
Loading
Loading