mirrors/ollama-for-amd
1
0
Fork 0
mirror of https://github.com/likelovewant/ollama-for-amd.git synced 2025-12-21 22:33:56 +00:00
Code Issues Packages Projects Releases Wiki Activity

Follow up to #10363 (#10647)

Browse Source 
The quantization PR didn't block all unsupported file types,
which this PR fixes.  It also updates the API docs to reflect
the now reduced set of supported types.
This commit is contained in:
Daniel Hiltgen
2025-05-12 15:23:31 -07:00
committed by GitHub
parent 0cefd46f23
commit 9d6df90805
4 changed files with 88 additions and 382 deletions
Download Patch File Download Diff File Expand all files Collapse all files

55
server/quantization.go
View File

@@ -70,23 +70,7 @@ func getTensorNewType(kv fsggml.KV, qs *quantizeState, newType fsggml.TensorType
newType = fsggml.TensorTypeQ6_K
}
} else if strings.Contains(name, "attn_v.weight") {
if ftype == fsggml.FileTypeQ2_K {
if kv.GQA() >= 4 {
newType = fsggml.TensorTypeQ4_K
} else {
newType = fsggml.TensorTypeQ3_K
}
} else if ftype == fsggml.FileTypeQ2_K_S && kv.GQA() >= 4 {
newType = fsggml.TensorTypeQ4_K
} else if ftype == fsggml.FileTypeQ3_K_M {
if qs.iAttnV < 2 {
newType = fsggml.TensorTypeQ5_K
} else {
newType = fsggml.TensorTypeQ4_K
}
} else if ftype == fsggml.FileTypeQ3_K_L {
newType = fsggml.TensorTypeQ5_K
} else if (ftype == fsggml.FileTypeQ4_K_M || ftype == fsggml.FileTypeQ5_K_M) &&
if (ftype == fsggml.FileTypeQ4_K_M) &&
useMoreBits(qs.iAttnV, qs.nAttnV) {
newType = fsggml.TensorTypeQ6_K
} else if ftype == fsggml.FileTypeQ4_K_S && qs.iAttnV < 4 {
@@ -114,54 +98,23 @@ func getTensorNewType(kv fsggml.KV, qs *quantizeState, newType fsggml.TensorType
} else if strings.Contains(name, "ffn_down") {
iLayer := qs.iFfnDown
n_layer := qs.nFfnDown
if ftype == fsggml.FileTypeQ2_K {
newType = fsggml.TensorTypeQ3_K
} else if ftype == fsggml.FileTypeQ2_K_S {
if iLayer < n_layer/8 {
newType = fsggml.TensorTypeQ4_K
}
} else if ftype == fsggml.FileTypeQ3_K_M {
if iLayer < n_layer/16 {
newType = fsggml.TensorTypeQ5_K
} else if useMoreBits(iLayer, n_layer) {
newType = fsggml.TensorTypeQ4_K
} else {
newType = fsggml.TensorTypeQ3_K
}
} else if ftype == fsggml.FileTypeQ3_K_L {
newType = fsggml.TensorTypeQ5_K
} else if ftype == fsggml.FileTypeQ4_K_M {
if ftype == fsggml.FileTypeQ4_K_M {
if useMoreBits(iLayer, n_layer) {
newType = fsggml.TensorTypeQ6_K
}
} else if ftype == fsggml.FileTypeQ5_K_M && useMoreBits(iLayer, n_layer) {
newType = fsggml.TensorTypeQ6_K
} else if ftype == fsggml.FileTypeQ4_K_S && iLayer < n_layer/8 {
newType = fsggml.TensorTypeQ5_K
}
qs.iFfnDown++
} else if strings.Contains(name, "attn_output.weight") {
if nExperts == 8 {
if ftype == fsggml.FileTypeQ2_K || ftype == fsggml.FileTypeQ3_K_S || ftype == fsggml.FileTypeQ3_K_M ||
ftype == fsggml.FileTypeQ4_K_S || ftype == fsggml.FileTypeQ4_K_M {
newType = fsggml.TensorTypeQ5_K
}
} else {
if ftype == fsggml.FileTypeQ2_K {
newType = fsggml.TensorTypeQ3_K
} else if ftype == fsggml.FileTypeQ3_K_M {
newType = fsggml.TensorTypeQ4_K
} else if ftype == fsggml.FileTypeQ3_K_L {
if ftype == fsggml.FileTypeQ4_K_S || ftype == fsggml.FileTypeQ4_K_M {
newType = fsggml.TensorTypeQ5_K
}
}
} else if strings.Contains(name, "attn_qkv.weight") {
if ftype == fsggml.FileTypeQ3_K_M || ftype == fsggml.FileTypeQ3_K_L {
newType = fsggml.TensorTypeQ4_K
} else if ftype == fsggml.FileTypeQ4_K_M {
if ftype == fsggml.FileTypeQ4_K_M {
newType = fsggml.TensorTypeQ5_K
} else if ftype == fsggml.FileTypeQ5_K_M {
newType = fsggml.TensorTypeQ6_K
}
}

223
server/quantization_test.go
View File

@@ -42,71 +42,6 @@ func TestGetTensorNewType(t *testing.T) {
ftype: fsggml.FileTypeF32,
expected: fsggml.TensorTypeQ6_K,
},
{
name: "attn_v.weight_q4_k",
kv: map[string]any{
"general.architecture": "foo",
"foo.attention.head_count": uint32(4),
"foo.attention.head_count_kv": uint32(1),
},
newType: fsggml.TensorTypeQ4_0,
tensor_name: "blk.0.attn_v.weight",
shape: []uint64{256},
ftype: fsggml.FileTypeQ2_K,
expected: fsggml.TensorTypeQ4_K,
},
{
name: "attn_v.weight_q3_k",
kv: map[string]any{},
newType: fsggml.TensorTypeQ4_0,
tensor_name: "blk.0.attn_v.weight",
shape: []uint64{256},
ftype: fsggml.FileTypeQ2_K,
expected: fsggml.TensorTypeQ3_K,
},
{
name: "attn_v.weight_q2_k_s_q4_k",
kv: map[string]any{
"general.architecture": "foo",
"foo.attention.head_count": uint32(4),
"foo.attention.head_count_kv": uint32(1),
},
newType: fsggml.TensorTypeQ4_0,
tensor_name: "blk.0.attn_v.weight",
shape: []uint64{256},
ftype: fsggml.FileTypeQ2_K_S,
expected: fsggml.TensorTypeQ4_K,
},
{
name: "attn_v.weight_q3_k_m",
kv: map[string]any{},
newType: fsggml.TensorTypeQ4_0,
tensor_name: "blk.0.attn_v.weight",
shape: []uint64{256},
ftype: fsggml.FileTypeQ3_K_M,
expected: fsggml.TensorTypeQ5_K,
},
{
name: "attn_v.weight_q3_k_m_i",
qs: quantizeState{
iAttnV: 2,
},
kv: map[string]any{},
newType: fsggml.TensorTypeQ4_0,
tensor_name: "blk.0.attn_v.weight",
shape: []uint64{256},
ftype: fsggml.FileTypeQ3_K_M,
expected: fsggml.TensorTypeQ4_K,
},
{
name: "attn_v.weight_q3_k_l",
kv: map[string]any{},
newType: fsggml.TensorTypeQ4_0,
tensor_name: "blk.0.attn_v.weight",
shape: []uint64{256},
ftype: fsggml.FileTypeQ3_K_L,
expected: fsggml.TensorTypeQ5_K,
},
{
name: "attn_v.weight_q4_k_m",
qs: quantizeState{
@@ -156,88 +91,6 @@ func TestGetTensorNewType(t *testing.T) {
ftype: fsggml.FileTypeF32,
expected: fsggml.TensorTypeQ8_0,
},
{
name: "ffn_down_q2_k",
qs: quantizeState{},
kv: map[string]any{},
newType: fsggml.TensorTypeQ4_0,
tensor_name: "ffn_down",
shape: []uint64{256},
ftype: fsggml.FileTypeQ2_K,
expected: fsggml.TensorTypeQ3_K,
},
{
name: "ffn_down_q2_k_s",
qs: quantizeState{},
kv: map[string]any{},
newType: fsggml.TensorTypeQ4_0,
tensor_name: "ffn_down",
shape: []uint64{256},
ftype: fsggml.FileTypeQ2_K_S,
expected: fsggml.TensorTypeQ4_0,
},
{
name: "ffn_down_q2_k_s_layers",
qs: quantizeState{
iFfnDown: 2,
nFfnDown: 3 * 8,
},
kv: map[string]any{},
newType: fsggml.TensorTypeQ4_0,
tensor_name: "ffn_down",
shape: []uint64{256},
ftype: fsggml.FileTypeQ2_K_S,
expected: fsggml.TensorTypeQ4_K,
},
{
name: "ffn_down_q3_k_m_base",
qs: quantizeState{
iFfnDown: 1,
nFfnDown: 8,
},
kv: map[string]any{},
newType: fsggml.TensorTypeQ4_0,
tensor_name: "ffn_down",
shape: []uint64{256},
ftype: fsggml.FileTypeQ3_K_M,
expected: fsggml.TensorTypeQ3_K,
},
{
name: "ffn_down_q3_k_m_16",
qs: quantizeState{
iFfnDown: 2,
nFfnDown: 3 * 16,
},
kv: map[string]any{},
newType: fsggml.TensorTypeQ4_0,
tensor_name: "ffn_down",
shape: []uint64{256},
ftype: fsggml.FileTypeQ3_K_M,
expected: fsggml.TensorTypeQ5_K,
},
{
name: "ffn_down_q3_k_m_8",
qs: quantizeState{
iFfnDown: 2,
nFfnDown: 3 * 8,
},
kv: map[string]any{},
newType: fsggml.TensorTypeQ4_0,
tensor_name: "ffn_down",
shape: []uint64{256},
ftype: fsggml.FileTypeQ3_K_M,
expected: fsggml.TensorTypeQ4_K,
},
{
name: "ffn_down_q3_k_l",
qs: quantizeState{},
kv: map[string]any{},
newType: fsggml.TensorTypeQ4_0,
tensor_name: "ffn_down",
shape: []uint64{256},
ftype: fsggml.FileTypeQ3_K_L,
expected: fsggml.TensorTypeQ5_K,
},
{
name: "ffn_down_q4_k_m",
qs: quantizeState{
@@ -264,19 +117,6 @@ func TestGetTensorNewType(t *testing.T) {
ftype: fsggml.FileTypeQ4_K_M,
expected: fsggml.TensorTypeQ6_K,
},
{
name: "ffn_down_q5_k_m",
qs: quantizeState{
iFfnDown: 2,
nFfnDown: 3 * 8,
},
kv: map[string]any{},
newType: fsggml.TensorTypeQ4_0,
tensor_name: "ffn_down",
shape: []uint64{256},
ftype: fsggml.FileTypeQ5_K_M,
expected: fsggml.TensorTypeQ6_K,
},
{
name: "ffn_down_q4_k_s",
qs: quantizeState{
@@ -290,59 +130,6 @@ func TestGetTensorNewType(t *testing.T) {
ftype: fsggml.FileTypeQ4_K_S,
expected: fsggml.TensorTypeQ5_K,
},
{
name: "attn_output.weight_8_expert",
qs: quantizeState{},
kv: map[string]any{
"general.architecture": "foo",
"foo.expert_count": uint32(8),
},
newType: fsggml.TensorTypeQ4_0,
tensor_name: "blk.0.attn_output.weight",
shape: []uint64{256},
ftype: fsggml.FileTypeQ2_K,
expected: fsggml.TensorTypeQ5_K,
},
{
name: "attn_output.weight_q2",
qs: quantizeState{},
kv: map[string]any{},
newType: fsggml.TensorTypeQ4_0,
tensor_name: "blk.0.attn_output.weight",
shape: []uint64{256},
ftype: fsggml.FileTypeQ2_K,
expected: fsggml.TensorTypeQ3_K,
},
{
name: "attn_output.weight_q3_k_m",
qs: quantizeState{},
kv: map[string]any{},
newType: fsggml.TensorTypeQ4_0,
tensor_name: "blk.0.attn_output.weight",
shape: []uint64{256},
ftype: fsggml.FileTypeQ3_K_M,
expected: fsggml.TensorTypeQ4_K,
},
{
name: "attn_output.weight_q3_k_l",
qs: quantizeState{},
kv: map[string]any{},
newType: fsggml.TensorTypeQ4_0,
tensor_name: "blk.0.attn_output.weight",
shape: []uint64{256},
ftype: fsggml.FileTypeQ3_K_L,
expected: fsggml.TensorTypeQ5_K,
},
{
name: "attn_qkv.weight_q3_k_m",
qs: quantizeState{},
kv: map[string]any{},
newType: fsggml.TensorTypeQ4_0,
tensor_name: "blk.0.attn_qkv.weight",
shape: []uint64{256},
ftype: fsggml.FileTypeQ3_K_M,
expected: fsggml.TensorTypeQ4_K,
},
{
name: "attn_qkv.weight_q4_k_m",
qs: quantizeState{},
@@ -353,16 +140,6 @@ func TestGetTensorNewType(t *testing.T) {
ftype: fsggml.FileTypeQ4_K_M,
expected: fsggml.TensorTypeQ5_K,
},
{
name: "attn_qkv.weight_q5_k_m",
qs: quantizeState{},
kv: map[string]any{},
newType: fsggml.TensorTypeQ4_0,
tensor_name: "blk.0.attn_qkv.weight",
shape: []uint64{256},
ftype: fsggml.FileTypeQ5_K_M,
expected: fsggml.TensorTypeQ6_K,
},
}
for _, tt := range cases {
t.Run(tt.name, func(t *testing.T) {