Place llama.cpp across several GPUs of a node with gpus_per_node

tp on a llama.cpp profile was a width nothing placed. gpus_per_node is
now the count: placement reserves that many cards on one node and the
engine splits layers across them. tp, pp and nodes are refused with
what to write instead.

Refs #2 (item 8).

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
This commit is contained in:
Ilker ManapandClaude Opus 5.5 committed 2026-10-01 11:56:34 +02:00
1 parent 057be16a92
commit cbd5521f4f
5 files changed
+99 -13

No files matched your search

+11
View File
@@ -47,6 +47,17 @@ as nvidia-smi does, by PCI bus, and CUDA's own order is fastest first — on a
node with two kinds of GPU, a shard could get a different card from the one it
was given.
### llama.cpp on several GPUs of one node
A llama.cpp profile takes several of a node's GPUs with `gpus_per_node: N`:
placement reserves N cards of its `gpu_class` on one node, the engine sees
those N (`CUDA_VISIBLE_DEVICES`) and spreads the layers over them
(`--split-mode layer`, in proportion to each card's free memory, which for
whole cards is even). `tp` or `pp` on a llama.cpp profile was a width nothing
placed — the replica stayed unplaced or ran on one card — and is now refused
when the model is added, saying to write `gpus_per_node` instead; so is
`nodes` above one.
## v0.10.6 — 2026-10-01
Published first as release candidates v0.10.6-rc.1 to rc.8; rc.5 failed its
+20 -13
View File
@@ -67,25 +67,25 @@ func (*LlamaCpp) ServesHTTP(ShardSpec) bool { return true }
// Validate checks the profile against the GPUs it would run on.
//
// The two refusals here are about not pretending: llama.cpp has no equivalent
// of a tensor-parallel group spanning machines, and its multi-GPU support is a
// different mechanism (--split-mode, --tensor-split) that this adapter does not
// drive. Accepting either would produce a replica that starts and then serves
// from one machine while the catalog believes it has several.
// llama.cpp spreads a model over a node's GPUs by layer (--split-mode layer),
// so gpus_per_node is how many it gets. It has no tensor parallelism and no
// group spanning machines, and accepting tp or nodes would produce a replica
// that serves from fewer cards, or one machine, while the catalog believes
// otherwise.
func (*LlamaCpp) Validate(p Profile, gpus []GPUInfo) error {
if err := validateCommon(p, gpus); err != nil {
return err
}
if p.Nodes > 1 {
return fmt.Errorf("profile %q: llamacpp does not span nodes; use vllm or sglang for a %d-node replica",
p.Name, p.Nodes)
}
if p.TP > 1 {
return fmt.Errorf("profile %q: llamacpp does not do tensor parallelism; "+
"it splits a model across local GPUs with --tensor-split, which this adapter does not drive",
p.Name)
if p.TP > 1 || p.PP > 1 {
return fmt.Errorf("profile %q: llamacpp splits a model across a node's GPUs by layer, not by tp or pp; "+
"write gpus_per_node: %d instead", p.Name, max(p.TP, 1)*max(p.PP, 1))
}
return nil
// The common checks count a replica's GPUs as tp; for llama.cpp the
// count is gpus_per_node.
q := p
q.TP = max(p.GPUsPerNode, 1)
return validateCommon(q, gpus)
}
// Commands returns the processes for one shard, which is always one process.
@@ -134,6 +134,13 @@ func (l *LlamaCpp) Commands(s ShardSpec) ([]Command, error) {
if len(s.GPUIndices) > 0 && !hasFlag(p.Args, "--n-gpu-layers", "-ngl") {
args = append(args, "--n-gpu-layers", "999")
}
// Several GPUs: the layers spread across all of them, which is the only
// split llama.cpp does without a matched interconnect. In proportion to
// each card's free memory, which for the whole cards placement gives is
// an even split.
if len(s.GPUIndices) > 1 && !hasFlag(p.Args, "--split-mode", "-sm") {
args = append(args, "--split-mode", "layer")
}
if len(s.GPUIndices) == 0 && !hasFlag(p.Args, "--threads", "-t") {
// What placement reserved first: it is what the node's other
// replicas were placed around.
+27
View File
@@ -334,3 +334,30 @@ func TestDevicesAreReadFromListDevices(t *testing.T) {
t.Errorf("no devices: %q", got)
}
}
func TestLlamaCppSpreadsAReplicaOverItsNodesGPUsByLayer(t *testing.T) {
four := Profile{Name: "t4x4", Engine: "llamacpp", GPUClass: "t4", TP: 1, GPUsPerNode: 4, ModelPath: "/m"}
if err := (&LlamaCpp{}).Validate(four, h100(4)); err != nil {
t.Fatalf("four GPUs on a four-GPU node: %v", err)
}
if err := (&LlamaCpp{}).Validate(four, h100(2)); err == nil {
t.Error("four GPUs were accepted on a node with two")
}
s := llamaSpec([]int{0, 1, 2, 3})
s.Profile.GPUsPerNode = 4
cmds, err := (&LlamaCpp{}).Commands(s)
if err != nil {
t.Fatal(err)
}
if v, _ := flagValue(cmds[0].Args, "--split-mode"); v != "layer" {
t.Errorf("--split-mode %q over four GPUs", v)
}
if !slices.Contains(cmds[0].Env, "CUDA_VISIBLE_DEVICES=0,1,2,3") {
t.Errorf("env %v: the replica must see all four of its GPUs", cmds[0].Env)
}
one, _ := (&LlamaCpp{}).Commands(llamaSpec([]int{2}))
if hasFlag(one[0].Args, "--split-mode") {
t.Error("a one-GPU replica was given a split mode")
}
}
+15
View File
@@ -381,6 +381,21 @@ func (m Model) validate() error {
}
continue
}
// llama.cpp on GPUs splits a model across one node's cards by layer,
// and how many is gpus_per_node. Not tp: llama.cpp has no tensor
// parallelism, and a tp it was given was a width nothing placed — the
// replica stayed unplaced, or ran on one card while four were meant.
if p.Engine == "llamacpp" {
switch {
case p.Nodes > 1:
return fmt.Errorf("model %q profile %q: llama.cpp does not span nodes; use vllm or sglang for a %d-node replica",
m.Name, p.Name, p.Nodes)
case p.TP > 1 || p.PP > 1:
return fmt.Errorf("model %q profile %q: llama.cpp splits a model across a node's GPUs by layer, not by tp or pp: "+
"write gpus_per_node: %d and leave tp and pp out", m.Name, p.Name, max(p.TP, 1)*max(p.PP, 1))
}
continue
}
// The arithmetic has to add up, or the engine is handed a shard spec it
// cannot satisfy and fails minutes later during model loading.
if want := p.GPUsPerNode * p.Nodes; want != p.TP*max(p.PP, 1) {
+26
View File
@@ -321,3 +321,29 @@ func TestAProfileCannotSetWhatRunsCodeInTheEngine(t *testing.T) {
t.Errorf("ordinary tuning refused: %v", err)
}
}
func TestALlamaCppReplicaOnSeveralGPUsSaysHowManyAsGPUsPerNode(t *testing.T) {
four := model("big", Profile{Name: "t4", Engine: "llamacpp", GPUClass: "t4", GPUsPerNode: 4})
c, err := New([]Model{four})
if err != nil {
t.Fatalf("gpus_per_node 4 on llama.cpp: %v", err)
}
m, _ := c.Resolve("big")
if pl := m.PlacementProfiles(); len(pl) != 1 || pl[0].GPUsPerNode != 4 {
t.Errorf("placement was not asked for four GPUs: %+v", pl)
}
// tp is a width llama.cpp has no way to use; refused, saying what to
// write instead.
for _, p := range []Profile{
{Name: "tp", Engine: "llamacpp", GPUClass: "t4", TP: 4},
{Name: "nodes", Engine: "llamacpp", GPUClass: "t4", Nodes: 2, GPUsPerNode: 1},
} {
_, err := New([]Model{model("x", p)})
if err == nil {
t.Errorf("%s: accepted", p.Name)
}
if p.TP == 4 && (err == nil || !strings.Contains(err.Error(), "gpus_per_node: 4")) {
t.Errorf("tp 4 should point at gpus_per_node: %v", err)
}
}
}