# DeepSeek-R1-Distill-Llama 70B on an L40S

Yes, quantized. Q4_K_M is the largest that fits, leaving room for 15k tokens of context.

Canonical page: https://makerportal.ai/lab/llm-vram/deepseek-r1-distill-llama-70b/l40s
Page title: DeepSeek-R1-Distill-Llama 70B VRAM on L40S — fits at Q4_K_M

This markdown document and the HTML page above are rendered from the same solved values at build time, by the same functions. Nothing here is written by a language model and nothing is fetched at request time.

## Key figures

- **Verdict:** Q4_K_M — 44.2 GiB usable of 48 GB
- **Usable memory:** 44.2 GiB — 92% of the card's 48 GB — an assumption for a dedicated card, not a single fraction applied to every device
- **Weights at FP16:** 131.4 GiB — 70.6B parameters
- **Weights (Q4_K_M):** 39.6 GiB — 131.4 GiB at FP16 · 70.6B params · 6 of 6 quant rows are measured files
- **KV cache:** 320.0 KiB per token at FP16 — 80 layers × 8 KV heads × 128 dimensions
- **Speed ceiling:** 19 tok/s — 864 GB/s ÷ bytes read per token
- **Largest context that fits:** 14,941 tokens at Q4_K_M
- **Trained context:** 131,072 — the cap the KV-cache figures are held to
- **Card bandwidth:** 864 GB/s — 48 GB GDDR6 ECC

## Every quant, against this card

Every quantization of DeepSeek-R1-Distill-Llama 70B against the 44.2 GiB usable on an L40S. "Measured" rows are the byte size of a real published file; "nominal" rows are the parameter count times the published bits-per-weight.

| Precision | Bits/weight | Weights | Source | Fits | Max context | Ceiling |
|---|---|---|---|---|---|---|
| FP16 / BF16 | 16.00 | 131.4 GiB | measured file | short by 87.3 GiB | — | 0.9 tok/s offloaded |
| Q8_0 | 8.50 | 69.8 GiB | measured file | short by 25.7 GiB | — | 2.6 tok/s offloaded |
| Q6_K | 6.56 | 53.9 GiB | measured file | short by 9.8 GiB | — | 5.0 tok/s offloaded |
| Q5_K_M | 5.66 | 46.5 GiB | measured file | short by 2.4 GiB | — | 8.9 tok/s offloaded |
| Q4_K_M | 4.82 | 39.6 GiB | measured file | yes | 14,941 | 19 tok/s |
| Q3_K_M | 3.89 | 31.9 GiB | measured file | yes | 40,127 | 23 tok/s |

## Questions this page answers

### How much VRAM does DeepSeek-R1-Distill-Llama 70B need?

131.4 GiB for the weights at FP16 — 70.6B parameters at two bytes each — and 39.6 GiB at Q4_K_M. The KV cache is on top of that and is not a fixed number: this model spends 320.0 KiB per token of context, so 8,192 tokens costs a further 2.5 GiB. An L40S makes 44.2 GiB of its 48 GB available on the assumption below.

### Can an L40S run DeepSeek-R1-Distill-Llama 70B?

Yes, quantized. Q4_K_M is the largest that fits, leaving room for 15k tokens of context. That is the weights and the KV cache together, against 44.2 GiB of usable memory.

### How fast will DeepSeek-R1-Distill-Llama 70B run on an L40S?

No faster than 19 tokens/second at Q4_K_M, and in practice below it. Decoding is memory-bound: every token reads all 39.6 GiB of weights plus the cache, and this card moves 864 GB/s. That division is the ceiling — no kernel, runtime or driver beats it, and a real runtime typically reaches 60–80% of it.

### Why does the context length change how much memory DeepSeek-R1-Distill-Llama 70B needs?

Because the KV cache holds one key and one value vector per token, per layer, for the whole conversation, and it is allocated separately from the weights. This model has 80 layers and 8 key/value heads of 128 dimensions, shared across 64 query heads — grouped-query attention, which divides the cache by 8. That works out at 320.0 KiB per token. Parameter count tells you nothing about this number.

## Method and limits

Weight bytes are the byte size of the real published file wherever one exists, and the model's exact parameter count times the published llama.cpp bits-per-weight where it does not. That distinction is on every row above and it matters at both ends: a sub-1B model's Q4_K_M file runs a third larger than the nominal figure because k-quants keep its embedding tables at higher precision, and an already-4-bit release cannot be quantized upward at all. The KV cache is 2 · layers · kv_heads · head_dim · bytes per token, summed over layers with each sliding-window layer capped at its window. The speed figure is a roofline bound, not a benchmark: bandwidth divided by bytes read per token, which no runtime exceeds and every runtime falls short of. The usable fraction of card memory is an assumption: 92% here, which is what this lane assumes for a dedicated card — the other class assumes 75%, so it is not one number applied to every device. That is the only assumed input on this page; every other figure is computed from the model config and the card's published specification. Nothing on this page is written by a language model.

## Related tool

[LLM VRAM & KV-Cache Footprint Calculator](https://makerportal.ai/lab/llm-vram-kvcache-calculator) — Interactive VRAM planner — change the model geometry, quantization, context length and card capacity and read the weight bytes, KV-cache bytes and the bandwidth ceiling on decode speed.

---

Source: MakerPortal — https://makerportal.ai/lab/llm-vram/deepseek-r1-distill-llama-70b/l40s. Free to quote and cite with attribution and a link to the canonical page.