Fixed save_imatrix to match old behaviour for MoE (#7099)

* Fixed save_imatrix to match old behaviour for MoE

This fix is simple and clear, but unnecessarily doubles the memory overhead..

* Fixed missing idx variable

* Unconditionally increment ncall

Co-authored-by: slaren <slarengh@gmail.com>

* Fixed 2 bugs in save_imatrix()

- Fixed segfault bug because the counts vector needed to be created.
- Fixed pre-existing bug didn't actually add to the counts for "--combine" option.

* ncall needs summing too

* Trailing whitespace

---------

Co-authored-by: slaren <slarengh@gmail.com>

common/sampling.cpp

@@ -35,6 +35,8 @@ struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_
 
     result->prev.resize(params.n_prev);
 
+    result->n_considered = 0;
+
     llama_sampling_set_rng_seed(result, params.seed);
 
     return result;
@@ -64,6 +66,7 @@ void llama_sampling_reset(llama_sampling_context * ctx) {
 
     std::fill(ctx->prev.begin(), ctx->prev.end(), 0);
     ctx->cur.clear();
+    ctx->n_considered = 0;
 }
 
 void llama_sampling_set_rng_seed(struct llama_sampling_context * ctx, uint32_t seed) {
@@ -253,6 +256,8 @@ static llama_token llama_sampling_sample_impl(
         }
     }
 
+    ctx_sampling->n_considered = cur_p.size;
+
     return id;
 }
 

common/sampling.h

@@ -81,6 +81,7 @@ struct llama_sampling_context {
     // TODO: replace with ring-buffer
     std::vector<llama_token>      prev;
     std::vector<llama_token_data> cur;
+    size_t n_considered;
 
     std::mt19937 rng;
 };

examples/imatrix/imatrix.cpp

@@ -19,6 +19,7 @@
 
 struct Stats {
     std::vector<float> values;
+    std::vector<int> counts;
     int ncall = 0;
 };
 
@@ -121,12 +122,10 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
         auto & e = m_stats[wname];
 
         ++e.ncall;
-        // NOTE: since we select top-k experts, the number of calls for the expert tensors will be k times larger
-        //       using the following line, we can correct for that if needed by replacing the line above with:
-        //if (idx == t->src[0]->ne[0] - 1) ++e.ncall;
 
         if (e.values.empty()) {
             e.values.resize(src1->ne[0]*n_as, 0);
+            e.counts.resize(src1->ne[0]*n_as, 0);
         }
         else if (e.values.size() != (size_t)src1->ne[0]*n_as) {
             fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]*n_as);
@@ -153,6 +152,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
 
                     for (int j = 0; j < (int)src1->ne[0]; ++j) {
                         e.values[e_start + j] += x[j]*x[j];
+                        e.counts[e_start + j]++;
                     }
                 }
             }
@@ -170,6 +170,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
         auto& e = m_stats[wname];
         if (e.values.empty()) {
             e.values.resize(src1->ne[0], 0);
+            e.counts.resize(src1->ne[0], 0);
         }
         else if (e.values.size() != (size_t)src1->ne[0]) {
             fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]);
@@ -183,6 +184,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
             const float * x = data + row * src1->ne[0];
             for (int j = 0; j < (int)src1->ne[0]; ++j) {
                 e.values[j] += x[j]*x[j];
+                e.counts[j]++;
             }
         }
         if (e.ncall > m_last_call) {
@@ -222,7 +224,13 @@ void IMatrixCollector::save_imatrix(const char * fname, const char * dataset) co
         out.write((const char *) &p.second.ncall, sizeof(p.second.ncall));
         int nval = p.second.values.size();
         out.write((const char *) &nval, sizeof(nval));
-        if (nval > 0) out.write((const char *) p.second.values.data(), nval * sizeof(float));
+        if (nval > 0) {
+            std::vector<float> tmp(nval);
+            for (int i = 0; i < nval; i++) {
+                tmp[i] = (p.second.values[i] / static_cast<float>(p.second.counts[i])) * static_cast<float>(p.second.ncall);
+            }
+            out.write((const char*)tmp.data(), nval*sizeof(float));
+        }
     }
 
     // Write the number of call the matrix was computed with
@@ -270,14 +278,28 @@ bool IMatrixCollector::load_imatrix(const char * imatrix_file, std::unordered_ma
             imatrix_data = {};
             return false;
         }
-        e.values.resize(nval);
-        in.read((char*)e.values.data(), nval*sizeof(float));
+
+        // When re-called from load_imatrix() with add set, this will already be created.
+        if (e.values.empty()) {
+            e.values.resize(nval, 0);
+            e.counts.resize(nval, 0);
+        }
+
+        std::vector<float> tmp(nval);
+        in.read((char*)tmp.data(), nval*sizeof(float));
         if (in.fail()) {
             printf("%s: failed reading data for entry %d\n",__func__,i);
             imatrix_data = {};
             return false;
         }
-        e.ncall = ncall;
+
+        // Recreate the state as expected by save_imatrix(), and corerct for weighted sum.
+        for (int i = 0; i < nval; i++) {
+            e.values[i] += tmp[i];
+            e.counts[i] += ncall;
+        }
+        e.ncall += ncall;
+
     }
     return true;
 }

examples/server/README.md

@@ -272,7 +272,7 @@ node index.js
 
     `logit_bias`: Modify the likelihood of a token appearing in the generated text completion. For example, use `"logit_bias": [[15043,1.0]]` to increase the likelihood of the token 'Hello', or `"logit_bias": [[15043,-1.0]]` to decrease its likelihood. Setting the value to false, `"logit_bias": [[15043,false]]` ensures that the token `Hello` is never produced. The tokens can also be represented as strings, e.g. `[["Hello, World!",-0.5]]` will reduce the likelihood of all the individual tokens that represent the string `Hello, World!`, just like the `presence_penalty` does. Default: `[]`
 
-    `n_probs`: If greater than 0, the response also contains the probabilities of top N tokens for each generated token. Default: `0`
+    `n_probs`: If greater than 0, the response also contains the probabilities of top N tokens for each generated token given the sampling settings. Note that for temperature < 0 the tokens are sampled greedily but token probabilities are still being calculated via a simple softmax of the logits without considering any other sampler settings. Default: `0`
 
     `min_keep`: If greater than 0, force samplers to return N possible tokens at minimum. Default: `0`
 

examples/server/server.cpp

@@ -2266,17 +2266,31 @@ struct server_context {
                 llama_token_data_array cur_p = { slot.ctx_sampling->cur.data(), slot.ctx_sampling->cur.size(), false };
                 result.tok = id;
 
-                const int32_t n_probs = slot.sparams.n_probs;
-                if (slot.sparams.temp <= 0 && n_probs > 0) {
-                    // for llama_sample_token_greedy we need to sort candidates
-                    llama_sample_softmax(ctx, &cur_p);
-                }
+                const size_t n_probs = std::min(cur_p.size, (size_t) slot.sparams.n_probs);
+                if (n_probs > 0) {
+                    const size_t n_considered = slot.ctx_sampling->n_considered;
 
-                for (size_t i = 0; i < std::min(cur_p.size, (size_t) n_probs); ++i) {
-                    result.probs.push_back({
-                        cur_p.data[i].id,
-                        cur_p.data[i].p
-                    });
+                    // Make sure at least n_probs top tokens are at the front of the vector:
+                    if (slot.sparams.temp == 0.0f && n_probs > n_considered) {
+                        llama_sample_top_k(ctx, &cur_p, n_probs, 0);
+                    }
+
+                    if (slot.sparams.temp == 0.0f) {
+                        // With greedy sampling the probabilities have possibly not been calculated.
+                        for (size_t i = 0; i < n_probs; ++i) {
+                            result.probs.push_back({
+                                cur_p.data[i].id,
+                                i == 0 ? 1.0f : 0.0f
+                            });
+                        }
+                    } else {
+                        for (size_t i = 0; i < n_probs; ++i) {
+                            result.probs.push_back({
+                                cur_p.data[i].id,
+                                i >= n_considered ? 0.0f : cur_p.data[i].p // Tokens filtered out due to e.g. top_k have 0 probability.
+                            });
+                        }
+                    }
                 }
 
                 if (!process_token(result, slot)) {