Llama on NVIDIA Ada

Can I run Llama 4 Scout 109B-A17B on an L40S?

Not at Q4_K_M — it needs 64.0 GB against 44.3 GB available. Drop to IQ3_XXS and it fits, at about 62.5 tokens per second.

Won't fit

64.0 GB of 44.3 GB · 144%
067 GB
Weights 61.7 GB
KV cache 1.5 GB
Runtime overhead 0.8 GB
Over the limit 19.7 GB

Short by 19.7 GB. You can run it with 32 of 48 layers on the L40S and the rest in system RAM, at roughly 9.40 tok/s — usable for batch work, painful for chat. A smaller quantisation or a shorter context is usually the better trade.

Generation9.40tok/s
Prompt processing2236tok/s
Max context0tokens
KV per 1K tokens0GB

Every quantisation of Llama 4 Scout 109B-A17B on a L40S

Highlighted row is the highest quality that still fits at 8K context.

QuantisationbpwWeights TotalVerdictMax ctxtok/s
FP16 / BF16 16.00 203.0 GB 205.4 GB Won't fit — 1.42
INT8 / W8A8 8.50 107.4 GB 109.7 GB Won't fit — 3.32
Q8_0 (GGUF) 8.50 107.4 GB 109.7 GB Won't fit — 3.32
FP8 (E4M3) 8.00 101.5 GB 103.9 GB Won't fit — 3.61
Q6_K 6.56 83.2 GB 85.6 GB Won't fit — 5.12
Q5_K_M 5.67 71.9 GB 74.3 GB Won't fit — 6.83
Q5_K_S 5.52 70.0 GB 72.4 GB Won't fit — 7.00
Q4_K_M 4.85 61.7 GB 64.0 GB Won't fit — 9.40
Q4_K_S 4.58 58.3 GB 60.7 GB Won't fit — 11.0
Q4_0 4.55 58.0 GB 60.3 GB Won't fit — 11.0
AWQ 4-bit 4.25 56.8 GB 59.1 GB Won't fit — 11.9
GPTQ 4-bit 4.25 56.8 GB 59.1 GB Won't fit — 11.9
MXFP4 4.25 56.8 GB 59.1 GB Won't fit — 11.9
IQ4_XS 4.25 54.2 GB 56.6 GB Won't fit — 14.0
Q3_K_M 3.91 50.0 GB 52.3 GB Won't fit — 18.6
IQ3_M 3.70 47.4 GB 49.7 GB Won't fit — 23.2
IQ3_XXS 3.06 39.4 GB 41.8 GB Fits, but tight 22K 62.5
Q2_K 2.63 34.1 GB 36.4 GB Runs comfortably 50K 70.8
IQ2_XXS 2.06 27.0 GB 29.3 GB Runs comfortably 88K 86.0
IQ1_M 1.75 23.1 GB 25.4 GB Runs comfortably 109K 97.3

Also worth checking

Llama 4 Scout 109B-A17B on every GPU
Full hardware tableEverything that runs on a L40S
Full model tableWhy the KV cache is the part that bites
Guide