boltbeam.fyi

Qwen3.8-27B W4A16-AWQ on RTX 4090

safetensors INT4_G128_ZP+BF16

Engines

vLLM

0.31.0 ↗measured 2026-11-09 19:41 UTCnewer release at the time: 0.33.0

54.4 tok/s

88% of the roofline61.7 tok/s possible

now, a token
18.37 ms
floor
16.20 ms
read, a token
15.59 GB
More
memory, measured
962 GB/s

What was run

python vllm_decode_drive.py --model Qwen3.8-27B-W4A16-AWQ --contexts 512 --batches 1 --tokens 64 --reps 1 --once 64 --max-num-batched-tokens 584 --model-context-limit 584 --capture-pass paired

Speed at context

128 tokens
vLLM55.6 tok/s
512 tokens
vLLM54.4 tok/s of 61.7
The pink tick is the roofline at that context.

Where the time goes

Now, a token
18.37 ms
Roofline floor, a token
16.20 ms
Gap to the floor
2.17 ms
Kernels a token
790

Per role ms a token. floor (bytes ÷ measured memory speed) above the floor

ffn_gate_up INT4_G128_ZP6.71 / 6.17 ms 92%
ffn_down INT4_G128_ZP3.34 / 3.08 ms 92%
ssm_out INT4_G128_ZP1.00 / 0.82 ms 81%
attn_gate + attn_qkv INT4_G128_ZP2.36 / 2.18 ms 92%
ssm_alpha_beta BF160.19 / 0.05 ms 26%
recurrent0.42 / 0.32 ms 76%
attention0.13 / 0.04 ms 30%
attn_kv + attn_qo INT4_G128_ZP0.71 / 0.63 ms 90%
attn_qo INT4_G128_ZP0.33 / 0.27 ms 81%
lm_head BF162.65 / 2.64 ms 100%
triton_red_fused__to_copy_add_fused_add…0.27 ms · no floor
kernel gaps0.08 ms · no floor
flash::flash_fwd_splitkv_combine_kernel0.06 ms · no floor
layer_norm_fwd_kernel0.06 ms · no floor
triton_poi_fused_mul_silu_slice_10.06 ms · no floor
25 other small kernels0.19 ms · no floor
Per kernel (34): ms now / ms at the roofline, launches a token
Marlin14.45 / 13.15 ms ×256
gemm · lm_head BF162.65 / 2.64 ms ×1
_causal_conv1d_update_kernel + fused_recurrent_gated_delta_rule_packed_decode_kernel0.42 / 0.32 ms ×96
triton_red_fused__to_copy_add_fused_add…0.27 ms · no floor ×128
gemm · ssm_alpha_beta BF160.19 / 0.05 ms ×48
flash_fwd_splitkv_kernel0.13 / 0.04 ms ×16
flash::flash_fwd_splitkv_combine_kernel0.06 ms · no floor ×16
layer_norm_fwd_kernel0.06 ms · no floor ×48
triton_poi_fused_mul_silu_slice_10.06 ms · no floor ×48
_fused_qk_rmsnorm_rope_gate_kernel0.03 ms · no floor ×16
vllm::reshape_and_cache_flash_kernel0.03 ms · no floor ×16
triton_poi_fused_zeros_30.02 ms · no floor ×32
_gumbel_sample_kernel0.02 ms · no floor ×1
triton_poi_fused_mul_silu_slice_20.02 ms · no floor ×16
triton_poi_fused_mul_sigmoid_view_00.02 ms · no floor ×16
triton_poi_fused_zeros_40.01 ms · no floor ×15
at::native::index_elementwise_kernel0.01 ms · no floor ×3
_bias_kernel0.01 ms · no floor ×1
_gather_block_tables_kernel0.00 ms · no floor ×1
at::native::reduce_kernel0.00 ms · no floor ×1
triton_red_fused__to_copy_add_embedding…0.00 ms · no floor ×1
_post_update_kernel0.00 ms · no floor ×1
_prepare_rope_positions_kernel0.00 ms · no floor ×1
_combine_sampled_and_draft_tokens_kernel0.00 ms · no floor ×1
_get_num_sampled_and_rejected_kernel0.00 ms · no floor ×1
at::native::vectorized_elementwise_kern…0.00 ms · no floor ×3
at::native::unrolled_elementwise_kernel0.00 ms · no floor ×1
_compute_slot_mappings_kernel0.00 ms · no floor ×1
at::native::_scatter_gather_elementwise…0.00 ms · no floor ×1
_prepare_pos_seq_lens_kernel0.00 ms · no floor ×1

No floor = the kernel reads no weights (norms, conversions, copies): all of its time is gap.

Run rtx4090-qwen3.8-27b-awq-vllm

Tested on

GPU
RTX 4090 · 24 GB
System RAM
189 GB
CPU threads
96
Memory, measured
962 GB/s
Temp, max
64 °C
Power
92 W mean of 450 W
SM clock
2562 MHz
Driver
595.58.03
Measured
2026-11-09 19:41 UTC · context 512 tokens · measured, in-model (nsys): split by role from the engine's own labels, 9 weight roles

Qwen3.8-27B on other GPUs